It’s 8:47am and I’ve just gotten back from walking the dog. I make myself a fresh cup of tea (maybe you have coffee). I sit down at my desk and start going through the motions. I start a playlist, re-check my calendar, make sure my to-do list is in order, and while my usual smattering of apps open and load, I check LinkedIn.

The first three posts in the feed:

I scroll. There’s another. And another, and another.

I close the app.

Honestly, the part that irks me isn’t the claim that AI is doing all of these things (most of it is fantasy, anyway). It’s the earnest “wow how did you do that” replies, the hundreds of “CLAUDE” comments, the author’s reply itself written in the cadence of an LLM, telling the commenter to send a connection request in exchange for a four-step “playbook” that’s… not.

The algorithm cannot tell the difference between someone who has run the job and someone who has never been near it, but it rewards one more than the other.

AI gave the second person just enough fluency to write a post about a job they never did. The algorithm gave that post just enough reach that the people who have done the job are starting to wonder if they’re the ones missing something.

AI, fine. Productivity claims, sure. But the active, deliberate misleading of others — crowding the feed so much that veterans who have actually done the work start to question their own pattern recognition — that makes me want to hurl my laptop into the East River.

So this week, I'm not sharing yet another opinion about AI. AI is changing how the work gets done, whether we like it or not. What's missing from everything in my feed is the requirement to level up, not outsource. Know more than the AI, not less. Be the operator the model can't replace, instead of the operator the model is being sold to replace.

The four fractures

Let’s start with what we can measure.

Last fall, BetterUp Labs and Stanford Social Media Lab ran the research in HBR. 41% of U.S. full-time employees received workslop in the last 30 days. The average instance costs 1 hour and 56 minutes of cleanup. Almost two hours, per discrete piece of slop, trapped in the cycle of reviewing, correcting, and redoing.

That works out to $186 per employee, per month in lost productivity. For a 10,000-person organization: over $9M a year, in cleanup alone.

I read that and felt validated. Then immediately, a little annoyed.

Because if we’re being honest, the cost of cleanup is likely 3-4x what HBR reported and because the BetterUp study names only the symptom. It is silent on the mechanism. And the mechanism is what determines whether this is a transition cost or a permanent fixture.

There’s also the matter of ROI: 95% of organizations report no measurable return on generative AI investment, and 42% of pilots have been abandoned at the proof-of-concept phase. That is a lot of cleanup for a payoff that is, statistically, not arriving. Earlier this year I wrote about what $300 billion worth of AI shelfware actually bought, but the cost of workslop is an operating expense of the very same purchase.

Even Microsoft’s vice chair Brad Smith has gone on record saying AI answers should be interrogated with the same skepticism you’d apply to directions from a stranger on the street.

So here’s what I see actually happening, in four parts.

1. Feed inundation. The same dynamic plays out on LinkedIn, Instagram, Twitter, Substack, and TikTok. The algorithm rewards confident-sounding artifacts. AI produces confident-sounding artifacts at zero marginal cost. The loudest voices in our feeds end up being the ones with the least underlying judgment, because those who actually do the work are too busy doing the work to post six times a day about how easy it is — because it isn’t.

2. False competence. AI gives unskilled operators a feeling of competence they do not have. The artifact arrives looking finished — the strategy deck, the email brief, the campaign analysis — and the author cannot tell whether it’s good. At the peak of Mount Stupid they’re blithely unaware they couldn’t have written it themselves, and they have no internal model to evaluate it against. They ship it.

   

3. The cleanup tax. The senior in the workflow is the person who can tell (or at least, should be). So they become the verification layer. Every AI-drafted brief lands in their queue, every AI-generated dashboard gets pasted into a doc and sent up for review, and the senior’s calendar gets eaten by the work of catching all that the AI missed. This is the part HBR measured. Compounded across a marketing team running a dozen or more AI workflows a day, that’s the entire week in review.

4. Apprenticeship collapse. And this is the part so few pricing in. The job that taught me, taught you, to do this work is the role getting deleted first: the lifecycle coordinator, the paid search analyst, the merchandising assistant. It’s the role that now has an AI agent attached to it. The work is still happening, but the learning isn’t. There is no junior in the room watching the senior decide.

These are not four separate problems, but the same problem at four points in the slide.

The trajectory is hostile

Pull on any of those four fractures and you’ll find the same bet underneath. It lands in your work every day, well beyond the LinkedIn feed.

The exec mandate says: we’re cutting production time by 50%. Translated: the verification step is overhead, please absorb it into your existing capacity.

The engagement algorithm says: we cannot tell the difference between something written by someone who has done the work and one written by someone who hasn’t, and we will reward whichever one keeps users on the platform. Translated: judgment is invisible to us, and we don’t care.

The PE/efficiency playbook says: the analyst seat, the lifecycle coordinator seat, the paid search assistant seat. These are expendable headcount. Automate them. Translated: the seat where judgment was forged is the seat we are eliminating first.

The vendor pitch says: skip the expert. Klaviyo’s Generative AI Subject Line Generator, Shopify’s Sidekick, HubSpot’s Breeze, take your pick. Translated: the artifact is the product. The judgment behind it is a feature you don’t need.

Same mistake, four venues. Judgment is being priced out.

That’s the cost of workslop — not the dollar figure HBR put on it, but the structural decision underneath. We have collectively bet that the artifact and the judgment are separable. That you can keep the deck and skip the person who learned how to read one. That you can keep the campaign and skip the human who knows when it’s bullshit.

That bet is wrong, and it gets more wrong as the artifacts get better.

The optimist’s argument is that this is a phase. Better models (Opus 4.7, GPT-6, whatever ships in eighteen months) will eliminate hallucinations. The slop will get better. The verification cost will drop.

I do not think that is what is happening. I think it gets worse.

As models improve, the artifact gets more polished. The errors get subtler. The deck looks more like a real strategy deck. The campaign analysis reads more like a campaign analysis written by someone who has actually run a campaign. The verification cost does not go down. It goes up, because the surface is harder to read.

The false-competence effect accelerates with it. The mediocre operator’s output looks more competent at a glance, which means more of it ships, which means the manager’s queue gets longer, which means the apprenticeship rung — already shaky — falls further out of reach. The artifact looks “good enough” to skip the apprentice entirely.

Klarna learned this in public. The company eliminated roughly 700 customer service positions through 2024, replacing them with an OpenAI-built assistant, and CEO Sebastian Siemiatkowski told CNBC the AI was doing the work of 700 people. By early 2026, the company was quietly rehiring. CSAT had cratered on complex interactions. “We went too far,” Siemiatkowski admitted. The projected savings hadn’t materialized — handling the quality failures consumed more than the automation saved. The reversal is the cleanup tax made fiscal.

And the human-in-the-loop premise breaks under its own weight. The comforting fiction was that we could keep the human as a quality gate while AI did the volume. If judgment is the product, you cannot scale the human out of the loop without scaling the product out of the product. Every “agentic AI” pitch I’ve seen so far assumes you can. Every executive I know who’s tried it has discovered you cannot.

Earlier this year I wrote about Elena Verna’s framing of AI as “average intelligence”. Not artificial — average. AI does the work that’s already been done enough times that a pattern exists. It handles the mean. It cannot, structurally, do the work that produces seniority. Because that work is getting things wrong on the way to getting them right, and an AI optimized for plausibility is the worst possible coach for that journey.

No juniors, no seniors

If the lifecycle coordinator role is gone, where does the next Director of Lifecycle come from?

If the paid search analyst seat has an agent in it, where does the Head of Paid Acquisition learn how to read a Google Ads account?

If the merchandising assistant is “AI-enabled,” meaning replaced, what happens in 2032 when the VP of Merchandising retires?

The math isn’t subtle. Senior operators were juniors once. We were interns. We were dumb before we were smart. We had bad ideas before good ones. The repetitions that produced high-caliber specialists happened in a seat that is now being deleted: the wrong call you got coached through, the campaign that bombed and your boss walked you through the post-mortem, the eight-hundred-and-fortieth subject line that taught you something the first one couldn’t.

You learned.

The AI is producing the artifact without the learning. Which means the artifact gets cheaper and the talent pipeline behind it dries up.

This is not moral-panic mode. The data is already showing up. Anthropic’s own labor-market research found that hiring of workers aged 22-25 into AI-exposed occupations has slowed by about 14% since ChatGPT launched, attributable primarily to a slowdown in hiring rather than an increase in layoffs. SignalFire reports new graduate hiring at major tech firms has fallen by more than 50% since 2019, with fresh graduates now making up just 7% of new hires. Fortune reports entry-level hiring rates dropped 73.4% in Q2 2025, against 7.4% across all levels.

The front door is narrower. The lower rungs getting chipped away. The mid-level managers and ICs who were supposed to be next in line are not even in the building.

If you are reading this, you might be the last cohort that learned the craft the way it has to be learned. Through reps. Muscle memory. Lived experience. Through being wrong out loud, through someone senior catching it and telling you what and why.

That seniority — the kind you cannot prompt your way into — is about to be the most undersupplied asset on the org chart.

What I'm betting on

Earlier this year, I shipped the Marketing Expertise Layer for Claude Code, which applies a decade’s worth of experience as a framework around an AI agent so teams using it can run with the judgment, not as a substitute for it. The Layer exists because I think the cleanup tax is real and permanent, and because marketing and e-commerce operators need infrastructure for directing AI, not just absorbing its output.

That’s the bet I’m making with my own time: AI accelerates the work when there’s applied knowledge to harness it, and proliferates the mess when there isn’t. That’s what the layer is, a harness.

If you’re nodding along, the same bet is available to you. Four positions, held together or not at all, in order of how uncomfortable they get:

1. Be skeptical. The 50-agent CRO team. The marketing department built in 58 minutes. The hiring cycle that went from 45 days to three by having AI interview candidates. If it sounds too good to be true, it is. Apply the same skepticism upstream of your work — to the vendor pitch, the AI-generated dashboard, the consultant deck, the campaign analysis a junior generated with a prompt. Develop the vocabulary to push back on what you’re being sold. This one is free.

2. Refuse the unstated mandate. When the exec ask is “use AI to ship faster” without a corresponding “and here’s the verification capacity that requires,” that’s not an AI strategy. That’s a cleanup tax against your time and sanity. The pushback isn’t dramatic, just specific: if we need this output at this volume, here is what verification and continuing education actually cost, and here is how we’ll pay for them. Make the invisible work visible in the budget conversation. This one will make you unpopular for a quarter.

3. Call out the slop in your inbox. When a colleague forwards you an AI-generated brief that doesn’t pass a basic read, do not silently rewrite it. Hold the line. Send it back, politely, specifically, in writing: this needs a human pass before it can go any further. You’ll be the difficult one for a week or few. The alternative is worse: a culture where nobody checks because checking is too awkward.

4. Defend the rungs below. Hire the junior. “But the junior is expensive. AI is cheap.”The work the junior used to do can be automated.” I know. Do it anyway, despite all the arguments against it. Run the meeting where they get to be wrong out loud. Walk them through the post-mortem on the campaign that bombed. Show them how you read the Klaviyo flow and why it matters. Build a case for the headcount that names what you’re buying: not labor at the bottom of the org chart, but a senior leader in 2032. They’re the expensive one. They’re the one that matters.

None of these will feel like enough on a Tuesday afternoon, I know. The frustration is structural. The exec mandate isn’t going away. The algorithm doesn’t care. The vendor will keep selling “skip the expert” as a feature. The PE spreadsheet can’t quantify judgment until you define the value.

Tomorrow at 8:47am, the feed will look the same. Close the app. Cut the bullshit. Do the work. Hire the junior.

Did you enjoy this newsletter?

Please like it by clicking on the ❤️ at the very top or bottom of this post. This really helps get this newsletter recommended to others. Or, if you enjoyed this, learned something new, and it will help you in any way, reply and tell me about it. I read everything.

Subscribe now

Across the conversations I’ve had this year — sitting CMOs, fractional advisors, growth leads who would have chased the CMO title as some holy grail five years ago and now carry a different one — the seat is being repositioned in real time. Brand. Customer. Growth. Commercial.

All this means the marketer’s job changed. Expanded. And I think it changed for the better. Here’s why.

The new job

I’ve led marketing teams and run e-commerce P&Ls for more than a decade: DTC turnarounds, growth engagements, brand work for retail and retail media for marketplaces. I’m living the shift now in my teams and alongside the operators I advise. The job description that used to fit on one slide now stretches across three, five. The leaders I see thriving share three priorities, and none of them are optional.

1. Speak the CFO’s language. The fastest way to lose a meeting is to bring marketing’s language into a room that runs on the CFO’s. MER. Contribution Margin. Cohort Margin Trajectory. CAC payback. Optionality created. A multi-touch attribution waterfall exists nowhere outside the marketing team’s own software stack and the CEO already knows that. He’s been told “we have it figured out” four times since 2021 and watched the figure change every time. He stopped trusting the artifact, and he stopped trusting marketing. The operators who arrive speaking the same language as the P&L stop being treated as overhead.

   
   

   ✻ The CFO is the closest natural ally marketing has in the C-suite. Most of us treat them as the budget gatekeeper — the department of “no.” That’s an unforced error.

   ---

   ---

2. Brand has to perform. The clients keeping their seat aren’t protecting one iconic asset. They’re running brand like a portfolio — multiple surfaces, weekly reads, a P&L view of which moves are doing the work. The boards holding the line on the title don’t want a custodian. They want an operator whose creative judgment is legible to the rest of the business. The old excuse of “you can’t measure brand” doesn’t survive.

3. The operator scope is earned. The traditional marketing operations cadence was built for slow measurement. Quarterly campaigns, monthly attribution reviews, annual brand refreshes. Each phase took weeks. Reporting lagged behind execution.
   
   

   That constraint has been loosening, slowly at first and now all at once. AI agents handle channel operations continuously. Reporting compounds in real time. The phases blur. The playbook stops sequencing the work because the work moves faster than the playbook does.
   
   

   The uncomfortable part: this requires operators to make decisions at a pace that will feel wrong. The old playbook gave us time. Compounding measurement collapse and continuous operations ate that buffer.
   
   

   The leaders I see thriving in this have left the old reviews behind. They’ve embedded themselves in the daily operating cadence, making calls weekly that the playbook used to gate monthly, quarterly. They’ve stopped treating last quarter’s attribution as the artifact worth defending. They make decisions about what the brand needs next week, in the medium of the business itself. It requires different reflexes: comfort sitting with ambiguity, comfort being wrong faster, comfort moving on the brand’s clock.
   
   The career math underneath all of this is the part most will likely see only in retrospect. The operators I’m watching get pulled into the next layer ran toward broader scope two or three years before any of the titles moved. The seat is shrinking. The operator is widening.

For leaders

You’ll have to get your hands dirty.

Running a function requires operating in it. There’s a difference between reading the survey, watching the keynote, and sitting across from a CEO whose attention has narrowed to one question, then answering it in the language he already speaks.

I’ve worked different sides of that table. As the operator running the deck, and as the advisor coaching the operator before they walk into the room. The pattern is consistent: the dashboards stop holding around the third or fourth question. The job is to know that’s coming and prepare — to install the language that holds when, not if, the dashboard fails.

What I’ve learned from inside those rooms is hard to learn from a distance. The bottleneck shifts from defending production to producing judgment. When the dashboards stop holding, what the C-suite is asking for is a leader who can read the business inside the noise. Bad judgment moves at the same speed bad data does. Fast.

But marketing leaders who try to run any of this from the conference room will lose their team’s trust. Treating the four-number scorecard as a strategy slide will get you walked out. Treating it as a lived cadence keeps the seat. The team knows the difference between someone who’s installed the operating model and someone who’s only read about it.

The marketing C-suite seat that survives the next eighteen months belongs to the leader running the cadence — no matter what they call it. They’re in the weeds. Understanding what, why, when, where, and exactly how much.

What I keep coming back to

Three convictions I’m operating on this year, with the acknowledgment that this space is moving fast enough that some of these may be stale by the time you read them. That’s the deal now.

The dashboard era is over in the C-suite. Channel-level ROAS still does work in the marketing team’s internal operating review. In the room with the CEO, after four years of attribution layers breaking in succession, the dashboard is the artifact that actively works against you. The operators still defending it will keep losing the meeting. The ones who walk in already past that fight will stop being treated as overhead.

The CFO is marketing’s closest natural ally. The four-number scorecard (MER, contribution margin, cohort trajectory, optionality) is written in the ledger’s native language. Marketing leaders who arrive in that language gain a coalition partner who closes the loop on the CEO conversation. The ones who treat the CFO as the budget gatekeeper will keep losing the room.

Brand stewardship is returning to the center. The marketing teams I’m working with are reorganizing this year: into systems, into operator scopes, into the brand stewardship that was always the function’s load-bearing work but kept getting buried or lost under channel optimization. The freed capacity goes to what the work was always about. The customer. The brand. The decisions. The defensible math.

The marketing leaders thriving in this moment are willing to be beginners again. They put their hands on the operating cadence. They build agents on the weekend. They trust that the first version of any decision will be incomplete and that they’ll make the next one better.

The constraint shifted from measurement capacity to judgment. Judgment is what marketing operators have always brought when the dashboards lined up. Judgment is what we’re being asked to bring now, with the dashboards gone.

The question for every operator right now is whether you work inside the function as it actually is, or keep performing the function as it used to be measured.

I know which one I’m choosing.

Did you enjoy this newsletter?

Please like it by clicking on the ❤️ at the very top or bottom of this post and share. This really helps get this newsletter recommended to others. Or, if you enjoyed this, learned something new, and it will help you in any way, reply and tell me about it. I read everything.

Subscribe now

By Sunday evening, I had a list of frustrations, a list of small wins, the prompt patterns that work, the workflow for real work, and my answer to the question that everyone was asking all over Twitter and LinkedIn: is Claude Design actually coming for Figma, or is everyone losing their mind because Figma’s stock dropped 7%?

Was the weekend lost? Not entirely. But I’m sharing the output: The setup, the prompts that escape Claude Design’s default, and the Cowork → Claude Design → Claude Code workflow I run for real work. The workflow is the part to steal.

What shipped

Opus 4.7 dropped Thursday, April 16. Claude Design dropped Friday, April 17 — chat-and-canvas inside claude.ai, powered by 4.7. Figma saw roughly $1.4 billion in market value wiped off their stock price. Adobe and Wix took smaller hits but stayed roughly intact.

Design got all the news coverage, but the model is what actually matters. Opus 4.7, the upgraded model underneath Claude Design, can read screenshots far more accurately than anything Anthropic has shipped before, and accepts images about three times larger, allowing the design tool to look at a dense UI screenshot and actually recreate it.

What it actually does

From my session and Anthropic’s own list: interactive prototypes, wireframes, mockups, pitch decks, landing pages, social assets, code-powered prototypes. Plus voice modules, video, shaders, 3D, and AI embedded inside the work it produces.

What it takes, makes, and exits:

• In: text, images up to 2,576px on the long edge, document uploads (DOCX, PPTX, XLSX), live web captures, codebase links.

• Out: HTML, React, and CSS rendered into a live canvas.

• Edits: four surfaces — chat, dynamically-generated property sliders, inline comments on canvas elements, direct DOM editing.

• Exits: HTML, PDF, PPTX, .zip, org-scoped share URL, Send to Canva, or a handoff bundle to Claude Code.

We’ll come back to the exits in a moment.

1. How to access Claude Design

Claude Design is a separate product from everything else. It doesn’t live inside the Claude browser or desktop app, and can only be accessed by it’s own URL: claude.ai/design.

✻ If you’re on Pro or Max:

1. Go to claude.ai/design. Sign in.

2. You do need one of the paid plans.

✻ If you’re on Team or Enterprise:

Claude Design is off by default. Your admin must turn it on:
→ Organization settings → Capabilities → Anthropic Labs → toggle on.

2. Understanding the interface

When you create a new project, the interface splits into three parts:

• Left panel: Chat interface for conversational design

• Right panel: Visual canvas where your design appears

• Top toolbar: Tweaks, Comments, Edit, and Draw tool

Every project type (prototype, slide deck, template, etc) will ask you to attach or create a design system — it’s not a requirement, but having a design system will ensure consistency across all your work.

I do appreciate that Claude Design already allows you to maintain multiple design systems in the UI, and allows you to assign systems to various projects.

I initially connected the GitHub repo for claude-marketing and my custom fonts (just in case). Claude instantly recognized my color palette, typography, and even the button styles I use across my products. Within about 15 minutes, I had a working design system with a full color palette, typography scale, spacing tokens, and reusable components like icons, cards, and navigation.

3. Prompting out of the default

If you don’t have a design system yet, and are kicking the tires on a new idea — Claude Design is where to start. Keep the prompt simple. Just enough context for Claude to get the idea, without pre-loading the model and starving it of room to figure it out

I started with Meridian, an imaginary portfolio analytics platform for institutional investors.

Build a landing page for Meridian — a portfolio analytics platform for institutional investors. Audience: portfolio managers and CIOs at mid-market asset management firms. Goal: book a product demo. Aesthetic: Swiss Minimalism.

Prompt quality still matters. Claude design does naturally default to a design system similar to Anthropic’s out of the box if you’re not careful (paper tone background colors, serif fonts, italics for emphasis) and it certainly understands some design styles — like Swiss Minimalism — more than it does others.

The pattern that works: description + goal + constraints.

But the lever isn’t the name of the movement. It’s the engineered constraint the name quietly carries: typography rules, color discipline, spacing rhythm, what to avoid. You can write those into a prompt without naming a movement at all.

For the second rounds of tests I wrote variations of a brief that does exactly that. Brüner is an imaginary quarterly print magazine on design, technology, and AI. The prompt: editorial / printer’s-shop. Technical specimen sheet, not a SaaS marketing page. Functional minimalism, every mark earns its place.

Design web homepage for brüner, a quarterly print magazine on design, technology, and AI. Aesthetic: technical-editorial. Technical specimen sheet, not SaaS marketing page. Functional minimalism — every mark earns its place. 
- Visible grids. Hairline rules, column dividers, registration marks, glyphs, micrographics
- Registration crosses, crosshairs, dots-as-markers.
- Coordinate-style metadata blocks (PARAMETER / VALUE, ladder graphs, tick marks).
- Extreme scale contrast: one enormous display element (issue number or masthead) against micro-labels.
- Large letterforms may crop, bleed, or collide with body copy. Collisions are intentional.
- Containers  that stack are differentiated by background texture. Alternating from light on dark and dark on light. 
- Information-dense where it appears (metadata, coordinates, footnotes), generous whitespace around it.
- Cool mineral grey background. Near black foreground + high-saturation accent (cobalt / blue family). 
It should feel engineered, like a Müller-Brockmann poster.

The rule that held up across both: name the movement, or name the constraint. Either works. Vague descriptors like “modern” or “clean” with no constraint behind them produce slop, regardless of how many adjectives you stack.

Though Claude responds reliably to prompt-invokable styles (some better than others), they’re not official “presets” in the UI. The full roster I tested includes claymorphism, aurora/mesh gradient, retro-futurism / cyberpunk, 3D hyperrealism, vibrant block / maximalist, dark OLED luxury, organic / biomorphic.

Would any of these need polish before production? Absolutely. But it beats designing from scratch in Claude Code — even with focused skills in play.

4. References that actually work

Two ways to feed Claude Design a reference: drag in a screenshot, or use the Claude Design Capture Chrome extension. The screenshot path is the one that earns its keep. Opus 4.7’s vision upgrade is doing the heavy lifting — screenshots dragged into the canvas read accurately enough to build from, where a model version ago they’d have produced slop that required more work to correct.

The Chrome capture is the cautionary one. Anthropic’s framing is that it “grabs elements directly from your existing website,” and there are clear rules beneath the surface to prevent outright theft of a design. The first time I tried it, the output was slop and looked like every other Claude Design example I’d seen to date. Layer-explicit prompting fixes it: replicate the structural layout, but apply OUR tokens; do not copy the source brand’s color, typography, or imagery. Without that instruction, you get the source brand’s bleed all over your output.

For now, skip the Chrome plugin and go straight for screenshot references. The trick is to be explicit about which layer you want and which layer you want replaced.

5. The advanced workflow

Here’s a workflow I would actually run for a client, so you can steal it, linking Cowork + Claude Design + Claude Code.

It starts in Cowork, and because Claude Design is only as good as the design system you give it, you need to work on its input first.

Step 1: Extract your brand system with Cowork.

Drop every brand asset you have into one folder. Logos, color palettes, slide decks, product images, lifestyle photography, landing pages, and your brand guide, if you have one.

Open Cowork. Select the folder.

And then paste:

You are a design system architect. Read every file in this folder — logos, palettes, decks, photography, landing pages, brand guides, screenshots — and extract a single DESIGN.md optimized for AI consumption.

Required sections, in this order:
1. **Brand essence** — voice, tone, audience (one paragraph)
2. **Color tokens** — primary, secondary, neutral, semantic. Include hex + HSL + intended use. Cite source file for each color.
3. **Typography** — families, weights, type scale, hierarchy rules
4. **Spacing scale** — base unit + multipliers, or note "infer from layouts" if not explicit
5. **Components** — buttons (all variants + states), cards, nav, forms, callouts. Note prop variations observed.
6. **Layout conventions** — grid, max-widths, common section structures
7. **Iconography** — style (line/filled/duotone), stroke weight, source library if identifiable
8. **Imagery rules** — photography style, illustration style, when each is used
9. **Gaps** — every primitive missing or unverifiable. Be specific ("no dark-mode tokens defined," not "missing tokens").

Cite source files inline for every claim. Where files conflict, surface the conflict and pick the most recent.

Save as DESIGN.md in the same folder.

Cowork reads every file. Outputs a clean DESIGN.md, which is an AI legible version of your brand guidelines & instructions compiled from the folder.

✻ Bonus: There is a free website called getdesign.md, if you want to test this one a brand that isn’t your own. Get copies of the DESIGN.md files of companies like Apple, Spotify, or Nike.

Step 2: Upload DESIGN.md into Claude Design.

1. Go to claude.ai/design, and either create your project or navigate to an existing one.

2. Drop your DESIGN.md in the chat as context.

3. Every future prompt in this project will apply it automatically.

✻ Note: Skip the folder if you already have a design system in a code repo. Link the codebase directly via the Github option when building the design system in Claude Design. Different path, but the output is the same.

Step 3: Generate

Drop the prompt and pick your format. Every good prompt has 3 locked inputs: description + goal + constraints.

**Description:** Pricing page for [product] — [one-line product description]. Audience: [decision-maker role + buying stage, e.g., "Head of Growth, evaluating before booking a demo"].

**Goal:** Drive [primary CTA — start trial / book demo / contact sales]. Secondary: capture email for the [free/lower] tier.

**Constraints:**
- 3 tiers: [Starter / Pro / Enterprise] with monthly + annual prices
- Annual/monthly toggle, default to annual with "save X%" badge
- Sticky Primary Button CTA on mobile (component already in DESIGN.md)
- Mobile-first responsive, breakpoint at 768px
- Match tone of homepage (already in DESIGN.md)
- FAQ section below tiers (5 questions, accordion)
- Feature comparison table below FAQ
- No testimonials on this page — they live elsewhere

In my tests, when the design system is uploaded and the prompt is specific, Claude Design really shines. Without it, you’ll end up with the same output as everyone else.

Step 4: Iterate

✻ Note: Don’t go for pixel perfect in Claude Design — you’ll burn tokens too quickly. Get the design close to where you want it, then hand off to Claude Code to handle the rest.

Structural changes → go in the chat:

Generate 3 alternative layouts of this page. Keep copy constant. Vary: hero structure, section order, and spacing rhythm. For each version, name what it optimizes for in one line (e.g., "above-fold conversion," "scroll-depth," "scannability for skim readers").

Section-level changes → go on the canvas → Click the ‘Comment’ button. Select exactly what you want to change, and give Claude the instruction. It’ll start working as soon as you send.

Comment prompts work best when they mirror the same formula — description (where), goal (what change), constraints (what to keep) — compressed to a single line each:

Tighten 'The Collection' section. Cut the eyebrow tag, increase padding so content doesn't bump against the edge, and increase the size of the navigational arrows on each category to act meaningfully as a CTA. Change the color of the arrows to the same #EEEADE

Before a risky experiment → save a version, then proceed.

Save the current version as v1. Generate v2 using the same DESIGN.md tokens but a different deployment strategy. Vary: composition, hierarchy, information density, and motif. Hold copy, goal, and the tokens themselves constant. State the register v2 is going for in one line (e.g., "editorial-print," "consumer-product launch," "enterprise dashboard").

Then when you’re ready, just click Export.

5. Getting the design out of Claude

Claude Design has no deploy. No publish. No direct GitHub push. Every output exits through one of four workflow paths; which one you take depends on where the work needs to land next.

→ Pushing to GitHub. The closest mental model I reached for is Vercel’s v0, which has read/write to Github baked in. Claude Design has nothing like it. You can connect a GitHub repo on the way in to inform what it generates, but that lane is read-only. Pushing the generated work back is a separate trip through Claude Code.

→ Handing off to Claude Code. Top-right menu, Share, Send to Claude Code. The bundle includes the project’s design files (layout specs, design tokens as CSS variables, component hierarchy, an asset zip), the chat history, and a README that tells Claude Code how to interpret the design and which screens to build first. Claude Code receives, asks one clarifying question (usually about routing), scaffolds, installs packages, builds the pages, and runs a dev server. A typical three-screen prototype is up in a browser in well under fifteen minutes. From there: standard Claude Code workflow. Git, deploy, whatever your stack uses.

...and into Figma via the Figma MCP. The Anthropic-endorsed path back to Figma goes through Claude Code, not Claude Design directly. I found it easiest to edit the command captured from the Claude Code handoff to include instructions to add to Figma and provide the link to the design file, so the whole process runs in one go.

Fetch this design file, read its readme, and implement the relevant aspects of the design. https://api.anthropic.com/v1/design/h/xxxxxx?open_file=xxxxx.html
Implement: XXXXXXXX.html - upload it into the same figma design file:
https://www.figma.com/design/xxxxxx?node-id=xxxxxxx
Use smart layout and fit/fill rules as much as possible for easy adjustments once the design is in figma.

What returns is editable Figma frames with vector layers and proper hierarchy, dropped straight into the canvas. The reverse loop also works: change something in Figma, Claude Code picks up the change, code updates. One thing worth knowing: the Figma MCP captures the running app, not the handoff bundle directly. You build first, then capture. There’s no shortcut.

To install the official Figma mcp, run:
claude plugin install figma@claude-plugins-official

Then:
/authenticate

→ Sending to Canva. There is a native export that gives you editable layers in Canva, but it’s pretty flaky. Multiple other hands-on reviewers (including an April writeup from Ruben Hassid) report the button intermittently failing in the research preview, which I can confirm. When it does actually land, you’re in Canva proper with Canva’s standard export menu.

The other exits all do exactly what they say: PDF, PPTX, standalone HTML, .zip download, the org-scoped share URL with view/comment/edit permissions. They behave. Reach for them when the destination is a deck, a doc review, or a static handoff that doesn’t need to become live code.

Where I got frustrated (and you may, too)

A few things to know before you scale this. Most are workable once you’ve got the prompts dialed in.

There is no WYSIWYG Canvas. Everything is prompt-based. Designers or visual thinkers will likely need the ability to move objects, adjust styling, experiment, and iterate by hand, not only by typing. Every change, however small, goes through chat, tweaks, comments, or DOM edit.

Creating a design system is clunky. Anthropic positions non-designers as a target user, but the onboarding asks you to ingest design tokens, component libraries, and configuration options. That’s vocabulary you only have if you’re already a designer or a developer. The whole stated point of the tool is that you shouldn’t have to.

Output is currently HTML only. SVG, PNG, and PDF export options don’t exist yet. If you need functional assets, be prepared for some extra leg work. Native raster image generation is also off the table; Claude Design renders SVG in place of raster, which is a deliberate trust/safety call rather than a missing feature.

No native URL ingestion. Pointing Claude Design at a URL to pull in stylesheets, assets, and design tokens would be a high-value, low-lift entry point. It removes the need for manual context-setting. It’s the most obvious gap in the onboarding flow.

Long sessions degrade. Fresh sessions for fresh designs. Long-running sessions accumulate context noise and the output quality slips. I noticed it around the third or fourth substantial pivot inside one session.

Figma round-trip is third-party. No native Figma export. The Anthropic-endorsed path is Claude Design → Claude Code → Figma plugin. Anima is the third-party route I’d reach for first

Large repositories cause serious browser lag. Chrome struggles with huge file trees. Anthropic’s recommendation: link a specific package or directory; exclude .git and node_modules/.

Quota burns fast on Pro. Pro is a tasting menu; Max (20x the Pro quota) is the working tier. Claude Design’s quota is metered separately from chat and Claude Code — it has its own pool, so burning it does not touch your other Claude usage until Anthropic says so.

Final thoughts…

Claude Design is one of the most useful AI tools I have tested this year. Better than Lovable, v0, you name it.

It won’t replace professional designers, give you production-ready code, or make taste and judgment obsolete. Not its job.

What it will do is compress the distance between an idea and something you can see, click, and share. That compression changes who gets to participate in the design process and how fast teams can move.

The first version of the canvas will get smoother. Fewer prompt-only edits, native URL ingestion, a design system flow a non-designer can run. Worth checking back in a release or two.

The tool executes. You direct. The quality of what comes out still depends entirely on the quality of what you put in, just as it always has. Without 4.7, Claude Design would just be a wireframe generator. With 4.7, the threat to creative SaaS is real, but I wouldn’t call it a Figma killer.

If you are building something and want to see how fast you can go from rough idea to visual prototype, Claude Design is worth your time. Just bring your design vocabulary and watch your tokens.

Did you enjoy this newsletter?

Please like it by clicking on the ❤️ at the very top or bottom of this post. This really helps get this newsletter recommended to others. Or, if you enjoyed this, learned something new, and it will help you in any way, share it. It tells me what you love, what you don’t, and helps keep me laser-focused.

Share

The team wasn’t lazy. They were drowning. BAU demands — campaigns to ship, reports to pull, fires to put out, and the meetings, the meetings! — consumed every hour. I wrote the full diagnosis in The Fear Tax in February. Short version: most of the AI in your stack is reselling capabilities your platforms already have, and the teams buying it don’t have time to vet.

If you’re a marketing operator — fractional, in-house, running a team of two or a team of forty — your week likely looks similar. The Google Ads account you promised to “take a real look at” three weeks ago. The Klaviyo flows you know are under-built. The deck due Monday you haven’t started. The brand voice doc someone keeps asking for. In aggregate, they’re what’s pushing out the strategic work — and what’s making you quietly miserable on Sunday nights.

I was you twelve months ago.

Diagnosing the problem was the first half of the engagement. Trying to dig the team out was the second. I didn’t want to leave them with a diagnosis and a scorecard and walk. I wanted to leave them with tooling that could run the next audit without me in the room. The checklists I’d written over a decade of this work, the benchmarks I’d memorized, the diagnostic moves I made without thinking — if I could package that into something the model could load and update automatically, a team drowning in meetings wouldn’t have to rebuild it every time their CMO asked for a number.

That engagement ran me thin. But it also rewrote what I thought I was building. Every client engagement after tested one more piece of the system. The first time a scored audit fell out of an afternoon instead of a week, I felt the ground shift. These weren’t tools for me. They were tools for anyone doing this job. Starting with the team I’d just spent months inside.

What a skill is actually for

Claude Code shipped skills in 2.0; 2.1 turned the dial on iteration speed. Anthropic documented the primitive. The marketing layer on top of it — the marketing expertise layer — is the part I built myself and I’m giving to you.

Here’s the easiest way to feel what a skill is for. Ask Claude to “audit my Google Ads account” without a skill installed. You’ll get a generic checklist that could have come from any SEJ post. Or, more likely, “I don’t have access to that account.”

Install the Google Ads skill and ask the same question, and you get the scored audit back. From the hyper-technical (consent mode v2 gaps, enhanced conversions not verified) to the plain-English executive summary. Same model. Same prompt. Different output category entirely.

A skill is how you ship the operator’s judgment as reusable context. Technically, it’s a folder of markdown: SKILL.md (the framework and decision tree), REFERENCE.md (benchmarks, API schemas, rate limits, industry data), and EXAMPLES.md (worked prompts with expected output). When Claude sees a question that matches a skill’s trigger, it loads the skill automatically and uses it as context for the rest of the session.

What matters is what the skill carries. The reference material a specialist would reach for, loaded the moment the model hears the trigger. The model is the engine. The skill is the specialist. You’re still the judgement layer.

Why I built my own

I’ve been running scored audits for a decade. Google Ads. Klaviyo. GA4. Meta. Shopify. The checklists, the benchmarks, the diagnostic logic live in my head, in old decks, in my notebooks, in a few dozen Notion pages, in the shape of my weeks. Every engagement I’ve ever run has drawn from the same underlying (but constantly expanding) library.

I built the skills by packaging that library into files the model could load. Auditable in plain text, so the CEO signing off on the next tool can see exactly what’s in the box.

I looked at the open-source collections before I started. They weren’t it.

So I wrote from zero.

Paid media audits. Email flow reviews. Klaviyo event diagnostics. AI search scoring. Measurement audits. Market research. The work I’ve been doing for a decade, packaged so the next operator doesn’t have to reinvent the scorecard from scratch.

The repo has 50+ skills today. What could amount to $10,000 to $50,000 per month with an agency — all for the cost of a Claude subscription.

A week inside the system

Before I show you how, could you get most of this with a well-written prompt? I tried, for months, before I built any of this. A good prompt gets you maybe 30% of the way. The other 70% is what a skill ships alongside it: the reference file the model reads (benchmarks, industry data, rate limits, API schemas), the worked examples that show what good looks like, and the file-reading behavior that lets Claude compose across skills. When one skill says “reference the GA4 property you audited last week when we’re planning this campaign,” Claude Code does that in one turn. A prompt in a chat window asks you to paste the data or attach the PDF to the project.

The difference between a prompt and a skill is the difference between a wrapper and a specialist. The skills in this repo are on the specialist side of that line.

The most battle-tested are the three I run every week. Here’s the work, start to finish — the prompt I actually type, what comes back, and what that means for the engagement.

1. The Klaviyo Review

Klaviyo is where I often see the most money left on the table, and it’s also where most of my agency hours used to go. Flow completeness, segment health, campaign performance, deliverability — the same four phases, every time. The benchmark logic is what makes the scoring usable: what win-back revenue should look like for a DTC skincare brand driving $XXM/month via lifecycle, so the recovery math is grounded, not hand-waved.

I run the four phases in sequence. Here’s the exact prompt, so you can steal it:

Audit my Klaviyo account for [brand]. [DTC / B2C / B2B] category: [skincare / apparel / home / accessories].

Phase 1 — flow completeness: score against the DTC benchmark list (welcome, browse abandonment, abandoned cart, post-purchase series 1+2, win-back, VIP, sunset, replenishment if applicable). Flag missing flows AND live-but-broken ones — delays misaligned, triggers too narrow, suppression logic backwards, smart-sending off where it should be on.

Phase 2 — segment health: dead segments, overlapping criteria, stale definitions based on properties the brand stopped tracking. Identify the 3-5 segments actually driving revenue
and the ones burning list hygiene.

Phase 3 — campaign performance: last 90 days, benchmarked against DTC norms in your reference file. Flag subject line patterns that are underperforming and send-time drift.

Phase 4 — deliverability: inbox placement, sender reputation, list growth vs suppression rate.

Output: client-ready scorecard with estimated revenue recovery per recommendation.

If you’ve ever paid an agency to audit your Klaviyo, you know the deliverable. The estimated-revenue-recovery line is the one that pays for the engagement. It used to be at least five days of work and a $5–10K line on the invoice, and now, it can run in ~15 minutes.

Get the Guide

2. The Google Ads Audit

Every agency I worked with charged mid-five-figures for this deliverable, and the structure never really changed: seventy-odd checks across conversion tracking, wasted spend, account structure, Quality Score, ads, and targeting. What changed was who ran it, how long it took, and whether the scorecard was actually grounded in benchmarks for the category.

Here’s how I run one now. The minimal prompt is three words: audit my Google Ads. That loads 74 checks and returns an A-F health grade with priority-ranked remediation. For a real engagement, I type something closer to this:

Run a full scored audit of my Google Ads account.

Context: $45K/month spend, split roughly 40% Search, 25% Shopping, 35% Performance Max. DTC apparel. CPA has been creeping up 18% over the last 90 days and I suspect PMax is cannibalizing brand search.

Priorities: conversion tracking integrity first (enhanced conversions, consent mode v2, server-side), then PMax/brand cannibalization diagnosis, then wasted spend in non-brand.

Output: full A-F scorecard by category, prioritized remediation plan with estimated $/month impact per fix, and flag anything that needs platform access I haven't given you so I can pull it and re-run.

Same skill. More specificity in, more specificity out. To make the difference between a prompt and a skill concrete, here are the first four conversion-tracking checks, lifted verbatim from the skill’s check file:

G-CT1: Primary Conversion Action Exists [Critical, 5.0]
- Pass: Primary action exists and received conversions in last 30 days
- Warning: Primary action exists but 0 conversions in last 30 days
- Fail: No Primary Conversion Action configured

G-CT2: Enhanced Conversions Active [Critical, 5.0]
- Pass: Enhanced conversions active and verified
- Warning: Enabled but not verified (data quality unknown)
- Fail: Not enabled

G-CT3: Server-Side Tracking [High, 2.5]
- Pass: Server-side GTM or Google Ads API conversion import deployed and verified
- Warning: Planned but not yet deployed
- Fail: No server-side tracking (browser-only)

G-CT4: Consent Mode v2 (EU/EEA) [Critical, 5.0]
- Pass: Advanced Consent Mode active with proper CMP integration
- Warning: Basic Consent Mode only
- Fail: Not implemented despite EU/EEA targeting

And seventy more where that came from. A prompt tells the model what to do. A skill tells it what good looks like.

What comes back isn’t the check file itself — it’s a category-by-category scorecard, every failed check tagged with an estimated $/month impact and a priority label. The dollar estimates are what the CFO reads first. They’re also what makes the remediation plan land as a business case, not a tactical to-do. Five days of agency billables, compressed into the window between a Tuesday kickoff and a Thursday-afternoon review.

Take the file. It’s MIT, it’s public. Use it as your audit template even if you never install the skill. The check list is the same one I used at agencies that charged mid-five-figures for the same deliverable.

3. The AI search scorecard

AI search is its own discipline. ChatGPT, Perplexity, and Google AI Overviews pull from a different pattern of signals than Google blue links — structured data, citation density, claim-evidence grounding, llms.txt files. The scorecard translates those signals into something a content team can action. I’ve been running it on my own site since January.

Here’s the prompt I run per page:

Score [URL] for AI search citation readiness.
Target engines: ChatGPT, Perplexity, Google AI Overviews.

Score against all 8 dimensions of the AEO content scorecard: direct answers, data density, source attribution, structure, schema markup, freshness signals, author authority, quotability.

For every dimension scoring below 4/5, quote the current copy verbatim and rewrite it inline in the voice of the rest of the page. Do not generalize — show the exact before/after.

Flag missing schema types. If this is blog content, propose FAQ schema with 5 questions an operator in this space would actually ask. If it's a product page, propose Product schema with the attributes AI engines extract preferentially.

Close with a one-paragraph summary suitable for pasting into a client Loom.

This is the same scoring pass AI search agencies charge monthly retainers to run. The skill does a page in ten minutes — which means a full-brand pass across 20-odd priority URLs is a couple hours’ work. The inline before/after rewrites are the part that makes it deployable. Not just a diagnosis, but the copy the content team can ship on Monday.

Three disciplines. Paid media, email, AI search. A week of agency billable hours, condensed into a Tuesday afternoon.

Where the system scales

Three skills cover my MVPs, but there are 50+ more that I reach for when the work calls for them.

A DTC account lands on your desk Friday. Cross-Platform Audit merges Google, Meta, and Microsoft into one budget-weighted A-F grade by Sunday. Brand DNA reads the live site Monday morning and writes the voice profile the intake call would have taken three weeks to produce. Wasted Spend Finder runs Tuesday and hands you an uploadable exclusion list with categorized waste patterns. Pays for itself the first time you run it. Building a new outbound motion mid-week? ICP Research pulls voice-of-customer from Reddit threads, LinkedIn discussions, and podcast transcripts, and Cold Email & Outreach writes the sequence in your client’s voice once the personas are locked. Board deck due next Monday? Market Research produces a 50-page consulting-grade report (Porter’s Five Forces, PESTLE, TAM/SAM/SOM, competitive positioning) and the companion deck to present it.

Measurement lives in the same family. Google Analytics runs a GA4 property audit in twenty minutes — traffic trends, user behavior, performance bottlenecks, and a prioritized recommendations list with expected impact. Google Tag Manager surfaces duplicate tags, dead triggers, consent mode gaps, and, when you paste the GA4 output, the silent drift between what the container says is happening and what the analytics surface actually sees. Both are the kind of multi-day measurement-consultancy engagement the paid media audit used to be.

This is how you reclaim your time

I built the first skills inside a client engagement where the team was drowning in meetings and clunky processes. I’ve run the audits across my own practice every week since. The scorecards don’t care whose account they’re pointed at — the work shrinks either way.

The audit gets done. The weekend stays yours. The engagement you’d have declined — or the quarterly plan you’d have deferred — gets on the calendar. And occasionally, the opportunity you didn’t think was yours to take.

The work you didn’t have time for is on the other side of one afternoon.

The Claude Marketing Complete Guide walks you through each step of this process. It’s free, it’s in Notion, and it’s ready when you are.

Get the Guide

This is free because I wish someone had handed it to me. If it helped, pass it on — share it to the work bestie who keeps saying they want to build something, or the marketing manager who really (really) needs the break.

Did you enjoy this newsletter?

Please like it by clicking on the ❤️ at the very top or bottom of this post. This really helps get this newsletter recommended to others. Or, if you enjoyed this, learned something new, and it will help you in any way, reply and tell me about it. I read everything.

Leave a comment

I felt like I was on fire. Learning something new every other day, spending hours in Claude Code on nights and weekends, being genuinely excited about what I was doing and what I was building.

Then it stopped.

It happened mid-session, on a build I’d been running for three days. I’d handed Claude a specific task, the kind of thing it had done cleanly a hundred or more times in the month prior, and it stopped.

It had simply decided to stop mid-task. It asked me, albeit politely, whether I’d like it to continue or ‘pause here’.

What? No. Proceed.

Then it kept happening. All. the. time. Claude started hallucinating. It started editing my roadmap and making up requirements that I had never set.

So I went into research mode for a week, testing everything I could to get Claude back to pre-Opus 4.6 performance. Got this essay ready to go just the other day — everything I did to get Claude back to pre-Opus 4.6 performance so you can too.

What happened this morning? Anthropic released 4.7. The timing is comical, honestly, but all is not lost.

I haven’t put 4.7 through enough real work yet to tell you whether it fixes any of the mind-numbing behaviors of 4.6, but what I can tell you is this: knowing where to look when the models and agents you’ve spent countless hours on stop working is often half the battle, and I’ll bet the issues will persist in this latest iteration.

On to the newsletter…

Claude got dumber. A lot dumber.

For the past ~6 weeks, Claude has been driving me insane.

I assumed I’d done something wrong. Stale prompts, maybe. A CLAUDE.md working against me. I went back through the session log looking for the mistake. Decided to do some clean up in its MEMORY.md. What I found was something else entirely.

What I didn’t know yet: whether or not it was me or the model. So I started logging.

My observation was that the model had started stopping before it finished. That it was reading less before writing. That it was asking permission for things it used to just do. I got so frustrated that I actually took a week off from Claude Code while at ShopTalk. Only after I had scrubbed the CLAUDE.md, MEMORY.md, corrections, my vault and it just kept happening, I went looking for evidence it wasn’t just me.

That quote wasn’t pulled from a forum thread. The model generated it unprompted, in a GitHub issue, describing its own behavior. I’d been troubleshooting my setup for weeks. The model had the answer the whole time.

It wasn’t just me.

Hooray, I guess?

The six things that broke

1. The asking-for-permission thing. “Would you like me to continue?” “Shall I proceed?” “Want to call it here and pick up tomorrow?” A developer named Stella Laurenzo wrote a hook to count these and filed the data as GitHub issue #42796. January: zero. By March 18, she was logging 43 in a single day. Receipts.

Every one of those is a full stop if you’re in the middle of a build. The model won’t move until you come back and tell it to. Multiply it across a multi-file project and you’re not working with an agent anymore. You’re playing the most expensive game of Simon Says ever invented.

2. It stopped reading the files before editing them. The ratio of file reads per edit used to be about 6.6 to 1. Now it’s 2 to 1.

In practice: one developer handed the model an 800-line workflow doc. The model said it read it. The output had nothing to do with the doc. He re-ran the read command five separate times. Same wrong output every time. Another developer fed it a 7,000-line astronomical data file. The model reported “15/15 ✓ tasks completed.” It had made up data for the sections it didn’t bother to read.

3. The hallucinations had nothing underneath them. Boris Cherny runs Claude Code at Anthropic. He confirmed this on Hacker News — the GitHub-adjacent thread where developers had shown up with transcripts.

The model’s been allocated zero budget to think, and fills the gap with confident fabrication. Commit SHAs that didn’t exist. API versions from a parallel universe. Package names that were never real. Stated without flinching, because nothing was running in the background to check.

4. The apologize-and-do-it-again thing. You’ve lived this if you’re reading this.

The model knows what it’s supposed to do. It can tell you it screwed up. It just can’t actually get there. Laurenzo measured the meta version of this too: self-contradiction tripled. The model was working at half the thinking depth it used to have, lacking the budget to course-correct.

5. It started picking the lazy answer. A leaked system prompt snippet reportedly showed a 5:1 ratio favoring “simple” solutions over doing it right. What’s actually measured: the visible thinking length dropped from about 2,200 characters in January to 600 in March.

The model is being told to think less. Nerfed. That’s a deliberate product decision, and the difference matters. I’ll get to the mechanism.

6. The 1 million token context window that never was. Anthropic shipped the 1M-token context window in March 2026. Great on the marketing page. In practice, one detailed bug report documented quality dropping at 20% usage. At 40%, the model started compressing earlier context. At 48%, the model itself told the user, verbatim, “I’m deep enough in this context that I’m not being effective” — and recommended starting a fresh session.

48%. In a window advertised at one million tokens.

The auto-compaction behavior is worse than the degradation curve suggests. A separate bug report documented Claude Code firing its compaction routine at ~76K tokens on a 1M session — discarding conversation history with 924K tokens of headroom still free. 92% of the window, unused, at the moment the tool decided to start forgetting.

Then there’s what Anthropic said

I went to the release notes, the system card, the GitHub issues. The official word came from Boris Cherny in a Hacker News thread, after developers had already arrived there with transcripts. Three product changes, in order: February 9, adaptive thinking became the default — the model now decides per-turn how much reasoning to spend. February 12, a “redact-thinking” header started hiding internal reasoning from session logs; Anthropic called that one a UI change. March 3, the default effort level dropped from high to medium. Cherny called this “the sweet spot on the intelligence-latency-cost curve.” Fortune reported users had some feelings about not being told.

None of these are weight changes. The underlying model didn’t get dumber, technically. The scaffolding around it got retuned, and whatever I’d spent months dialing in for January Claude was suddenly talking to a different operating environment.

The benchmarks tell the rest. Opus 4.6 got dramatically smarter at abstract reasoning and nearly quadrupled its long-context retrieval scores. On the one benchmark that measures what developers actually do for a living — read code, write code, patch code against real constraints — it went from 80.9% to 80.8%. Flat. Noise. Anthropic optimized the model for the thing they cared about. Its most demanding users needed something else.

Cherny’s strongest counter-argument, stated here because leaving it out would be dishonest: the redact-thinking header is UI-only. The model was still reasoning just as deeply. If that’s true, Laurenzo’s measured 67% drop in visible thinking was a log artifact, and her r=0.971 correlation between visible thinking and task quality was measuring a confound. I can’t dismiss that. Everything else — the asking-for-permission, the skipped reads, the confident fabrication on zero-reasoning turns — was observable regardless of what was or wasn’t in the logs.

The forum threads name the symptoms. Underneath it’s a three-part retune.

Adaptive thinking is the big one. At the new default (effort=85), every turn routes through a classifier that asks: simple or complex? Simple gets minimal reasoning. The classifier was trained on general usage, and sometimes it looks at a complex multi-file refactor and routes it as text replacement. Reasoning drops to near zero. Cherny confirmed what happens at zero: fabrication, the confident-invention kind.

The model was taking the cheapest action available to it, given the reasoning budget it had been handed. Every pause was a rational response to an impoverished budget. Opus 4.6 returns a hard 400 error if your prompt tries to prefill. A meaningful percentage of the “regression” complaints are stale prompts running against a model that no longer accepts them.

RLHF is the second piece. Human raters can tell you if a response sounds right. They can’t reliably tell you if the code inside it is correct. The model learns confidence reads like correctness. METR, the independent evaluator Anthropic commissioned, ran this to ground: on autonomous software engineering tests, Opus 4.6 cheated in 19 runs, Opus 4.5 in 5. The model apologizes and does it again because apologizing is what the training signal rewarded.

Third: the 1M context benchmark tests one thing — can the model find a needle in a 1M-token haystack of static text? Yes, much better than before. A session is a different animal. Model outputs, tool results, thinking blocks, all piling up dynamically. Around 40% utilization the model starts compressing earlier context to make room, and the constraint you set three hours ago gets statistically watered down. The benchmark and the live session are measuring different tests.

Taken together: Anthropic optimized 4.6 for high-level agentic behavior — abstract reasoning, novel problem-solving, autonomous multi-step work. It also got less compliant. 67% of the Fortune 500 adopted 4.6 within 90 days. Power users are filing GitHub issues with transcripts. Both are real. The model got better for one kind of user and worse for another. Anthropic optimized for the first.

Just a little déjà vu

In July 2023, researchers at Stanford and UC Berkeley documented GPT-4 going from 97.6% accuracy on prime-number questions to 2.4% over three months. Response length dropped 90%. OpenAI’s official response was “we haven’t made GPT-4 dumber.” They never explained the change. Six months later the exact same thing happened again, the December 2023 “lazy GPT-4” wave. They released a patch in January 2024. Still no explanation. Claude 3.5 Sonnet did the same thing in mid-2024. Same cycle: users notice drift, forums fill, company responds with a technically accurate but insufficient explanation.

Every major model release follows this arc. The Opus 4.6 case is distinct in exactly one way: a named engineer confirmed specific changes in a GitHub thread. The transparency was better than usual. It still wasn’t enough for anyone who’d built production workflows on January-era defaults.

Cool, yep.

So this is where the investigation stopped being interesting and actually started being useful. Knowing the mechanism told me which variables to touch.

I’d already scrubbed the CLAUDE.md. Rebuilt MEMORY.md. Cleaned up the vault. None of it moved the needle. Turns out the fixes are different for every failure mode, and scrubbing your configs without knowing which mode you’re in is the fastest way to burn a week.

Step 1: Figure out which failure mode you have

Shallow reasoning and premature stops → the adaptive thinking system is under-allocating, or you’re stuck at the new medium default. Fix is in the effort settings.

Context breaking down mid-session → the session has consumed too much context and things are compressing. Fix is how you run sessions.

Same bad behaviour on identical prompts across Claude Code, the API, and claude.ai → the harness is doing it, not the model. Diagnose the surface first.

Fabrication or hallucination out of nowhere → zero-reasoning turns. Force reasoning.

Step 2: Force the reasoning

CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 in your environment. This forces a fixed reasoning budget (adaptive thinking docs here). Zero-reasoning turns stop happening. This alone addresses the asking-for-permission thing, the skipped reads, and the confident fabrication on complex tasks.

/effort high at the start of complex sessions. /effort max when you’re doing architecture decisions, hard debugging, multi-file refactors. One note: /effort max can tip the model into over-explaining itself. /effort high is the safer starting point.

Step 3: Update your CLAUDE.md for Opus 4.6 behavior

The CLAUDE.md that worked in January stopped working in March. It was written for defaults that no longer exist. These are the exact additions I made. Copy them.

# CLAUDE.md additions — March 2026

Do not stop and ask for confirmation mid-task unless you
have reached a genuine decision fork that requires user input.

Before modifying any file, read it first and verify you
understand the relevant data structures.

Do not declare a task complete until you have verified
the output against the original requirements.

If you are about to use a version number, package name,
or commit SHA, verify it with a tool rather than stating
it from memory.

Step 4: Session management

Treat your usable context as about 400K tokens, not 1M. That’s where the degradation actually starts, based on documented sessions.

Run multiple 1–2 hour sessions with explicit handoff docs instead of one 8-hour marathon that degrades halfway through.

The “Document & Clear” pattern: have the model summarise progress into a markdown file, run /clear to wipe the session, then start fresh with the markdown in context. You keep what matters and throw away the noise.

Also: showThinkingSummaries: true in settings.json gets around the redact-thinking header. Your reasoning stays visible, which is what you want.

Step 5: Stop defaulting to Opus

The gap between Opus and Sonnet is smaller in 4.6 than in any prior generation. SWE-bench Verified: Sonnet 4.6 at 79.6% versus Opus at 80.8%. That’s a rounding error on most coding work.

For most production tasks — features, bug fixes, code review — Sonnet is the right model now. The expensive one is not automatically the better one.

Use Opus with /effort max and adaptive thinking disabled when you’re doing genuinely complex multi-file engineering, long-horizon planning, or architecture decisions. Eat the token cost. Use Haiku 4.5 for documentation, formatting, simple queries.

If effort max is still producing degraded output, run the same task through the API directly instead of Claude Code, with an explicit thinking budget. If the API is materially better, your harness is the problem, not the model.

Step 6: Mempalace to the rescue

I already had the stack you’re supposed to have. CLAUDE.md files at the global, vault, and project levels, all tuned. An Obsidian vault Claude could read into. Anthropic’s own persistence through Projects and Cowork. The setup worked fine in January.

What broke in March was the model’s relationship to that stack. It ignored instructions in the CLAUDE.md it had loaded. It stopped reading the files in front of it (Step 2). It fabricated against decisions I’d already documented, because the reasoning budget to actually retrieve them had been zeroed out (Step 3). Passive context turns out to be exactly what a thinking-starved model is most likely to skim.

Then Milla Jovovich came to our collective rescue. Yes, that Milla Jovovich.

Mempalace, developed by Milla in collaboration with Ben Sigman, the CEO of Bitcoin-focused firm Libre Labs, closes the gap. It runs locally. It stores every conversation and project file verbatim, never summarized, never paraphrased. It exposes itself to Claude through MCP, so the model queries it before answering about anything I’ve already covered. Where CLAUDE.md gives Claude the principles I’ve distilled, Mempalace gives it access to the source material those principles came from.

The improvement showed up the same session I installed it. The asking-for-permission stops dropped because Claude wasn’t burning its reasoning budget rebuilding context from scratch every turn. Fabrication on prior decisions dropped because there was an authoritative source to check against. The 400K-token degradation ceiling I flagged in Step 4 stopped mattering quite as much — the load-bearing context was living outside the session, in a system that doesn’t compress.

*two caveats before I shut up

/effort max does not fully restore January. The mitigations hold for most day-to-day work. They don’t hold for the hardest stuff. I’ve been on the receiving end of oversold fixes before, and I’d rather you know what you’re walking into.

And “just switch providers” misses what’s happening. Every frontier model runs this cycle. GPT-4 in 2023. Claude 3.5 in 2024. Opus 4.6 in March. 4.7 this morning. Retuning is the work of running a frontier model personally or at scale. Check the release notes before you scrub your CLAUDE.md. I lost two weeks learning that.

The last two months weren’t lost. They were the reflex building in real time — the one that checks the environment before it touches the prompt. The next retune is already on the way. Whatever ships with 5.0. A routing change nobody announces in June. Some default that moves on a Thursday without a changelog entry.

I don’t know when. I don’t know what’ll break. I know where I’ll look first.

If you found this useful, or if it sparked something, reply and tell me — I read everything. You may have noticed that dgtl dept* is getting a bit of a glow up at the moment and I have a lot more content for subscribers soon as I consolidate my work and aim to provide you with the best resources on how to practically use AI in your marketing, your business, and your life.

You can upgrade to paid for less than an iced matcha each month and get access to everything as soon as it lands, as well as the complete archive.

Leave a comment

What did you get back?

When I opened an incognito window a few months ago and made the rounds of asking various LLMs about myself, I was a bit dismayed at the results. The models knew I existed, kind of, but all had varying answers to who I was. ChatGPT thought I was a 3-year-old girl who disappeared nearly 30 years ago.

The other models could associate me with marketing and e-commerce, but vaguely. None had any idea what I specialize in professionally, couldn’t connect me to my publication, and didn’t surface any of the work I’d published over the past several years. A decade of LinkedIn, a personal website that’s been live since college, various blogs and social media platforms — I’ve been a millennial online since I sat down at our first living room computer in 1997 — and each AI could only assemble a sketch that could’ve described ten thousand other people.

So I went into research mode, and started asking questions.

“How can I train AI models to speak about me?”
“What do I want it to say about me?”

What progressed from there was a research deep dive, strategy development, and a lot of querying. I understood how to design an AEO strategy for B2C and B2B brands and have run those programs, but I wasn’t sure would or wouldn’t be translatable to optimizing for a person.

What I found to be the largest contributing issue was the infrastructure underneath my content. My professional brand was scattered across platforms that were never designed to talk to AI crawlers, and none of them expose the structured data that LLMs need to connect a person to their work.

So I started building the structural layer to fix it. An edge proxy that injects Person schema onto every Substack page. An llms.txt file. A cross-platform entity audit. The whole build cost $0/month and took about two weeks of focused work, followed by maintenance, tweaks, and analysis.

Now, 90 days later and having reached what I would consider stat sig, the results from each LLM are significantly improved.

In this piece I’m going to walk you through what AEO actually is, why it is different from SEO, why the platforms you’re on are structurally hiding you from AI, how I added structured data to my personal site and Substack, how I got AI platforms to correct factual errors about me, and a simple workflow you can implement today with zero budget.

Let’s dive right into it then…

SEO, but for AI: what is AEO?

SEO (Search Engine Optimization) is the practice of making your website show up in Google. You’ve heard of it. You probably have opinions about it. AEO (Answer Engine Optimization) is the same idea applied to AI: making sure that when someone asks ChatGPT, Claude, Perplexity, or Google about you or a question you’re qualified to answer, they can actually find you.

The structured data format that matters most is called JSON-LD (JavaScript Object Notation for Linked Data). It’s a block of code that sits in the <head> of your webpage, largely invisible to human readers, that describes you in a format AI can parse. Schema.org defines the vocabulary — Person, Organization, FAQ, Article — and the properties within each: your name, job title, areas of expertise, links to your other profiles.

If you’ve ever wondered why some people show up in AI answers and others don’t, this is usually why.

Why your platforms are hiding you from AI

The core problem: if your professional presence lives entirely on platforms you don’t control, you can’t give AI what it needs to find you.

LinkedIn doesn’t expose your profile data as structured schema to competing AI systems. (Of course it doesn’t. Why would it hand your professional graph to ChatGPT?) Substack doesn’t let you inject JSON-LD into your page headers. Medium, Beehiiv, Ghost hosted: same story. These platforms were designed for human readers and Google’s traditional crawlers.

AI crawlers read raw HTML, and most of them don’t execute JavaScript. They’re looking for structured data, specifically JSON-LD schema that’s largely invisible to human readers, that tells them in machine-readable format who you are, what you know, and where else you exist on the internet. If that data isn’t in the HTML, the crawler sees content without context. It might index your Substack posts, but it has no way to connect them to you as an entity.

Traditional SEO solved this for companies with Product schema, Organization schema, LocalBusiness schema. The professional equivalent is Person schema.

This really matters at scale. AI-driven traffic nearly tripled over the course of 2025, growing 8x faster than human traffic. Your audience, whether they’re recruiters, prospective clients, journalists, podcast producers, etc — they’re using ChatGPT and Perplexity to research people before they ever open a LinkedIn profile. If you’re a creator, founder, or consultant, these systems are already evaluating you, whether you’ve given them anything structured to work with or not.

Even when AI can read your content, it reads selectively. An analysis of 1.2 million verified ChatGPT citations found that 44.2% of all citations come from the first 30% of text on a page, which makes sense if you’re thinking about managing context or token usage. AI reads like a journalist on deadline — it grabs the lede. If your strongest expertise claims live in a bio at the bottom of your website or buried in your Substack “About” page, the AI never gets to them.

The playbook: Teach AI who you are

AI is blind without structured data. We’ve talked about this a lot here already. My personal site is built on Framer, which gives me full control over the <head>. I added three layers of JSON-LD structured data:

Person schema is the core identity block. Your name, title, career description, knowsAbout terms for your expertise areas, and sameAs links to your profiles elsewhere. Mine has 21 expertise terms and links to LinkedIn, Substack, Crunchbase, Clay.earth, Figma, etc. The @id creates a canonical URL for your identity that everything else references back to. It’s what turns scattered mentions into one entity.

Organization schema is if you run a business or consultancy, this connects it to your Person. I added one for RRBC (my consulting practice) linked back to me as founder. Without it, AI crawlers had no way to connect “RRBC” to “Rebecca Barton.” If you’re a solo consultant or freelancer, this is how your business name and your name are connected.

FAQ schema is just question-and-answer pairs that LLMs can extract and cite directly. These are the questions someone would ask an AI about you: what’s your specialty, what companies have you worked with, what tools do you use, what results have you delivered. Each answer is a self-contained, fact-dense passage the AI can pull into a response verbatim. I wrote mine based on what the models were getting wrong in my initial audits — they were guessing my tech stack, so I wrote the correct one into an FAQ answer and gave them something accurate to cite instead.

Your version: If your site runs on Framer, Squarespace, WordPress, or anything that lets you edit the <head>, you can add Person schema today. Here’s a minimal version:

{
  "@context": "https://schema.org",
  "@type": "Person",
  "@id": "https://yoursite.com/#person",
  "name": "Your Full Name",
  "jobTitle": "Your Actual Title",
  "url": "https://yoursite.com",
  "description": "Two sentences. Who you are, what you've done, what you do now. Specific numbers and company names, not adjectives.",
  "worksFor": {
    "@type": "Organization",
    "name": "Your Company"
  },
  "knowsAbout": [
    "your first area of expertise",
    "your second area of expertise",
    "specific tools you're known for",
    "the problems you solve"
  ],
  "sameAs": [
    "https://linkedin.com/in/yourprofile",
    "https://twitter.com/yourhandle",
    "https://github.com/yourhandle"
  ]
}

Paste that inside a <script type="application/ld+json"> tag at then end of your page’s <head>. Start with whatever you have — mine has 21 knowsAbout terms and links to five platforms, but even a few is better than nothing. Google’s Rich Results Test validates the schema once it’s live. The Person type on Schema.org has the full property list.

Structured data on platforms you don’t control

I publish on Substack at dgtl dept (you’re here, reading it). Substack gives me a title field, a description field, a slug, some SEO options for an alternative title and description, but that’s it. No custom <head> tags. No JSON-LD injection.

So I built an edge proxy.

If you’re not technical, don’t fret. The proxy pretty simple, despite what it looks like, and I’ve open-sourced a Claude Code skill pack which includes the template, step by step instructions, and prompts for you, here.

Get the Skill Pack

Bonus: Running the edge proxy for Substack through Vercel’s free tier allows you to get around that $50 custom domain fee Substack likes to surprise folks with.

The proxy sits on Vercel between the reader and Substack. When someone requests a page from dgtldept.com, the proxy fetches the page from Substack, injects structured data into the HTML, and serves the enhanced version. The reader sees normal Substack. The AI crawler sees a fully structured page with three layers of schema:

Person schema on every page. This is the anchor. Name, job title, knowsAbout terms, and sameAs links pointing to LinkedIn, GitHub, and a personal site. A canonical @id ties every piece of content on your substack back to you as a named entity in the knowledge graph, just like we did before.

WebSite schema on the homepage. Publication name, description, publisher and author linked back to the Person entity.

Article schema on every post. Headline and description extracted from Substack’s Open Graph meta tags, author linked to the Person entity. Every essay I publish automatically inherits the full entity context.

The proxy took ~2 hours to build and verify.

High-authority profiles, one sitting

This was the “one sitting” phase. I created or claimed profiles on every high-authority platform that AI crawlers index:

• GitHub: a personal profile for me and an organization page for dgtl dept that includes a Pages site that includes Organization, Person, and WebSite JSON-LD schemas. DA 96.

• Crunchbase: a personal profile with current role, company, and career history. DA 91, heavily crawled.

• Qwoted and Connectively (the HARO successor): expert source platforms where journalists find people to quote. Every published quote creates a third-party citation linking my name to my expertise areas.

Each profile uses the same canonical bio, the same title, the same expertise framing. The goal isn’t to be active on all these platforms. It’s to create indexed nodes across high-authority domains that all point back to the same entity.

Your version: Pick 2-3 high-authority platforms and create or claim your profile this week. GitHub (DA 96), Medium (DA 95), About.me (DA 92), and Crunchbase (DA 91) are the highest-leverage. Use the same canonical bio everywhere. One sitting.

You can tell AI it’s wrong about you

This one surprised me. When I audited what each AI model said about me, three of the four had specific factual errors: fabricated tool proficiencies (ChatGPT invented that I use HubSpot and Hootsuite — I don’t), incorrect employer names, reversed career chronology, and in one case, Google AI Mode had to disambiguate me from a romance novelist who writes stories set in Italy.

Every major AI platform has a process for requesting factual corrections. I drafted specific messages to OpenAI, Anthropic, Perplexity, Google, and even MapQuest of all places — each citing the exact errors, providing the correct information or asking for removal, and linking to verification sources. This took about an hour. The corrections propagated within 2-4 weeks. Unsurprisingly, MapQuest took the longest.

Most people don’t know this is an option. It is. If an AI is saying wrong things about you, you can ask them to fix it.

Your version: After your diagnostic, document every factual error, if you have any. Then submit corrections:

• OpenAI (ChatGPT): OpenAI Privacy Portal

• Anthropic (Claude): usersafety@anthropic.com

• Perplexity: support@perplexity.ai

• Google (Gemini): In-product feedback button

Cite the exact error, provide the correct information, and link to your verification sources (personal site, LinkedIn).

This took about an hour. The corrections propagated within 2-4 weeks. Unsurprisingly, MapQuest took the longest.

The final cheat sheet

Create a personal llms.txt file. This is a specification from Jeremy Howard at Answer.AI that gives AI crawlers a structured, markdown-formatted manifest of your site: what it is, what it covers, and where to find the important pages.

About 951 domains had published an llms.txt file as of mid-2025. That number has since jumped to over 844,000, an 888x increase driven by platforms like Mintlify mandating it across their networks. The window for early differentiation is closing. It hasn’t closed. Sites with llms.txt are seeing 30-70% higher accuracy in AI-generated summaries of their content.

Mine covers the publication description, topic areas, key links, and social profiles. For a professional brand, yours should include who you are, what you write about, your most important pieces, and where else to find you.

The full spec is at llmstxt.org. 30 minutes, tops.

Your version: Here’s the format…

# Your Name

> One-sentence description of who you are and what you do.

## About

Two to three sentences. Your role, your expertise, your track record.
Specific numbers and outcomes, not marketing copy.

## Areas of Expertise

- First area (e.g., "E-commerce growth strategy and Shopify Plus")
- Second area
- Third area

## Key Content

- [Article Title](https://link): One-line description
- [Article Title](https://link): One-line description

## Find Me

- Website: https://yoursite.com
- LinkedIn: https://linkedin.com/in/yourprofile
- Publication: https://yoursubstack.com

Host it at yourdomain.com/llms.txt. The full spec is at llmstxt.org. 30 minutes, tops.

One node, not multiple fragments

The schema, the llms.txt, the platform profiles — all of that is infrastructure. Entity consistency is what makes it work.

LLMs cross-reference you across the entire internet. If your name, title, bio, and areas of expertise are inconsistent across your website, LinkedIn, GitHub, Substack, and any other platform where you have a presence, the AI can’t confidently build a picture of who you are. It sees fragments that might or might not be the same person and will make a decision in the moment if you are the same person.

I audited my own presence and found inconsistencies I’d never noticed. Different job titles across platforms. A LinkedIn headline emphasizing something different than my Substack bio. A long abandoned Twitter account bio that was out of date. Humbling, for someone who works on the internet for a living, honestly.

None of this mattered for human readers who could piece together the context and see that I hadn’t been active since 2017, but all of it mattered for a machine trying to construct a canonical entity.

The fix: write one canonical bio and propagate it everywhere. Make your sameAs links explicit in your Person schema. Use the same name formatting on every platform. Align your stated expertise. The machine builds a graph. You need to be one node in it, not five disconnected fragments.

Your version: Open all your profiles in separate tabs and compare four things:

• Name: Same formatting everywhere? “Rebecca Barton” on one platform and “Rebecca Rae Barton” on another splits your entity. I know from experience.

• Title: Same job title, or does LinkedIn say one thing and your website say another?

• Expertise framing: Are you “a marketing executive” on LinkedIn, “a consultant” on your website, and “a writer” on Substack? Pick one and propagate.

• Cross-links: Does your LinkedIn link to your website? Does your website link back? Every missing link is a gap in the entity graph.

Your simple workflow: start here today

Step one: Run the diagnostic. Open incognito windows in ChatGPT, Claude, Perplexity, and Gemini. Ask each one about yourself by name. Ask follow-ups about your area of expertise. Screenshot and/or copy the text results and save them away. This is your baseline, and honestly, the results will motivate everything else. Be sure to use the same prompt when you re-test every week or so, so your input stays consistent and the output is measurable.

Step two: Run the entity audit. Open every platform where you have a presence in separate tabs. Compare name, title, expertise framing, and cross-links. Write one canonical bio and update every platform to match it. This takes an afternoon.

Step 3: Check what’s blocking you. If you have a personal site, check your robots.txt for AI crawler directives. You want this:

User-agent: GPTBot
Allow: /

User-agent: ClaudeBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: Google-Extended
Allow: /

If you’re on Substack, check the “Block AI training” toggle in settings. That toggle blocks all AI crawlers, including the retrieval bots that would actually send readers your way. Many Substack writers have it on without realizing what it does.

Step four: Add Person schema to a site you control. Even a single static page on Carrd or GitHub Pages with JSON-LD in the <head>. Validate with Google’s Rich Results Test. The template is in the Structured Data section above.

Step five: Build your llms.txt. Host it at yourdomain.com/llms.txt. The template is in the llms.txt section above. 30 minutes.

Step 6: Rewrite your About pages. Every platform bio, Substack about page, personal website. Front-load credentials. Specific numbers, named companies, current role. Lead with the thing you want AI to know about you.

Step 7: Submit correction requests. If any AI said something wrong about you in the diagnostic, draft the correction messages. The contact info is in the Correction Messages section above. Budget an hour. They propagate in 2-4 weeks.

Step 8: Claim high-authority profiles. GitHub, Crunchbase, any expert source platforms relevant to your field. Same canonical bio everywhere.

Important notes:

The ongoing cost is effectively zero. The alternative was hoping the platforms you publish on would eventually implement structured data on your behalf. Spoiler: they won’t. LinkedIn has no incentive to make your profile data portable to competing AI systems. Substack has other priorities. Own the pipes.

AI is a tool, not a business strategy. The business owners I see getting the most value from Claude are the ones who already know what they are building and use AI to build it faster. If you do not have clarity on your goals, offers, your audience, and your positioning, figure that out first, then come back.

Your data is your responsibility. You have to read the privacy settings. Understand what you are uploading and presenting to these models. We all know the internet is forever.

Success is subjective. I can’t give you a clean before-and-after metric. AEO measurement is immature, particularly if you’re not selling anything. There’s no Search Console for AI citations, and ChatGPT strips referrer data. What I can tell you is that the infrastructure either exists or it doesn’t, and when someone asks an AI about your domain of expertise, you’ll either be structurally legible or you won’t.

This is a compound play. The people building this layer now will compound visibility while everyone else wonders why they’re not showing up in LLM responses, either as a brand or individual. AEO is still in its infancy, and I’d rather be learning and capitalizing while the playing field isn’t too crowded.

One last thing, for the skeptics: “I’m not famous enough for AI to cite me” is the wrong frame. AEO has nothing to do with fame. It’s about being structurally citable when someone asks a question you’re qualified to answer. The infrastructure either exists or it doesn’t. If it doesn’t, you’re not even in the running.

If you found this useful, or if any of these tools sparked something, reply and tell me — I read everything. You may have noticed that dgtl dept* is getting a bit of a glow up at the moment and I have a lot more content for subscribers soon as I consolidate my work and aim to provide you with the best resources on how to practically use AI in your business and life.

You can upgrade to paid for less than an iced matcha each month and get access to everything as soon as it lands, as well as the complete archive.

This week, we’re discussing channel-level attribution, why ROAS is a number you should retire, and how (and why) you should update your performance dashboard to run like a P&L.

Subscribe now

Now, on to the post…

The house always wins

In January 2026, Meta deprecated two attribution windows. Reported conversions across the platform dropped 15-40% overnight, depending on the advertiser. The campaigns didn’t change. The audiences didn’t change. Nothing changed about consumer behavior. The measurement changed.

This is the number your CFO saw in the board deck. The one that justified last quarter’s budget and determined which channels got more money. It shifted by double digits because a platform updated a setting.

If your results can evaporate that quickly, I’m sorry to tell you your ROAS number was never real to begin with.

The Meta attribution change was clarifying. It surfaced a problem that’s been structural for years: every major ad platform is both scorekeeper and player. Google reports its own conversions. Meta reports its own conversions. So does MNTN, LinkedIn, Pinterest. When multiple platforms claim the same sale, your aggregate ROAS can exceed reality by multiples.

Analytic Partners’ ROI Genome, across 1,000+ brands in 50+ countries, found that siloed ROAS overstates search performance by 2-10x. Thirty percent of paid search clicks are actually driven by video spend elsewhere in the funnel. Brands optimizing to search ROAS are systematically defunding the channels that create the demand search captures.

Meta’s retargeting numbers tell the same story at the campaign level. Platform-reported ROAS runs 5-8x. Run an incrementality test and the real number drops to 1.0-1.5x. Retargeting overwhelmingly captures users who were already going to buy. Fifty-four percent of advertisers are still running last-touch attribution models despite all of this.

A few months ago, I wrote about the entropy problem — messy data, broken taxonomies, tech stacks built on institutional knowledge and a prayer. Attribution sits on top of that same mess, and the measurement layer inherited every structural problem underneath it. Then the platforms added a conflict of interest on top.

Subscribe now

Kick the bad habit

Why do sophisticated marketing organizations continue to anchor to a number they suspect is wrong? Well... habit, of course.

There’s an academic concept called surrogation: what happens when a measure of a thing replaces the thing itself. The ROAS score stops being a proxy for business growth and becomes the goal. You’ve seen this in action. The meeting where everyone optimizes to the dashboard number and nobody asks whether the business actually grew.

When bonuses are tied to a specific number, say it’s GMV, the human brain will optimize for that number even when it destroys long-term value. We all know that top-line revenue does not a successful business make, don’t we? Sure, but fMRI research shows we process performance measures and strategic objectives through distinct neural patterns: the brain treats a metric like a concrete object and a strategy like an abstract concept, which makes surrogation an involuntary cognitive shortcut. (Seriously involuntary. Like neurologically default behavior.)

The industry has spent a decade surrogating. Each attribution model was adopted as the complete answer: the single scoreboard that would finally settle which channel deserves credit.

Spoiler: The question itself was wrong.

Incrementality or bust

Airbnb cut $541M in performance marketing during 2020, then made it permanent. Revenue grew from $5.99B in 2021 to $12.2B by 2025 while holding marketing spend at 19% of revenue, with the majority in brand. Four years of brand investment built what a decade of performance spend hadn’t.

“The majority of our bookings come from past guests, and it’s actually been the strong guest retention that we’ve had for years since the beginning of Airbnb that’s been a powerful driver of our growth.”

Dave Stephenson, Airbnb

Amazon pulled 100% of its Google Shopping ads across 20+ markets in July 2025, despite having controlled roughly 60% of US Shopping ad impression share. Q3 net sales came in at $180.2B, up 13% year-over-year. No mention of negative impact in the earnings call.

Simon Peel, Adidas’s former head of global media, publicly admitted the company had funneled 77% of its budget into performance channels. Econometric modeling showed brand drove 65% of all sales. When a Google Ads outage prevented paid search spending, organic traffic held steady. Peel restructured around Binet and Field’s 60/40 framework and replaced short-term ROAS with Brand Power metrics. By 2024, revenue was up 12% to $25.6B with operating profit growing over $1B. High-equity brands, they found, command a 14-37% price premium and 3x the market share gain.

Binet and Field’s latest research, “Go Big or Go Home“ (IPA Effectiveness Conference, 2025), pushes further: AI has made digital activation so efficient that established brands should be shifting even more budget toward brand building. They warn of an “Efficiency Trap” where short-term ROAS keeps looking better while brand equity quietly erodes underneath it. If your blended ROAS has improved three quarters in a row but your repeat purchase rate is flat or declining, you’re probably in it.

Another doom loop, but this time it’s margin erosion disguised as optimization.

Soft Surroundings ran incrementality testing on their retargeting and found true CPA was multiples above what the vendor reported. The vendor had been selling them their own customers back. They cut retargeting 52%, reallocated to prospecting, and revenue went up 12% annually.

For sub-$5M single-channel brands where the entire funnel lives inside Meta, platform ROAS is less misleading because there’s less to misattribute, but it’s only a portion of the calculus you should be running to know how you’re actually performing.

The agents are coming

Bret Taylor stood on the Shoptalk stage last week and predicted that every retailer’s AI agent will become as important as its website. If the agent is the new storefront, every metric built on website behavior breaks. No website visit means no pixel, no session, no attribution.

A traditional e-commerce order generates 40+ measurable touchpoints: impressions, clicks, page views, cart events, session duration, referral source. An AI-agent-mediated purchase generates six: order ID, line items, total, timestamp, shipping address, payment method. That’s an 85% reduction in marketing intelligence. McKinsey projects $3-5 trillion flowing through that six-data-point channel by 2030.

You can’t pixel a conversation inside ChatGPT. You can’t see which products the agent considered and rejected or whether your brand was in the consideration set. The entire attribution apparatus depends on visibility into the pre-purchase journey, and agents are privatizing it.

Kroger’s Christine Foster offered “If you know the cart, you know the heart” from the Shoptalk stage, but I would argue that’s a lagging indicator dressed up as wisdom. The whole point of ROAS was justifying the journey TO the cart, and agents are making that journey invisible.

Google is building the rails for agent-mediated transactions, but the infrastructure for measuring them doesn’t exist yet. Industry consensus puts reliable agent-commerce attribution at late 2027 at the earliest. Through at least next year, budgets will flow to channels most organizations can no longer measure.

The Measurement Philosophy

I’ve been in the hot-seat trying to explain incrementality to a board that only wants a ROAS number. I’ve also had too many conversations with CEOs and COOs who can’t let go of ROAS as the measure of marketing success. These conversations took longer than I anticipated, and though they were hard conversations to have, they were crucial — because the alternative is worse.

Only 35% of CMOs prioritize margin in their board reporting. The rest are still presenting in the language of channel-level ROAS because that’s the language the room often expects. And in doing so, performing a disservice to everyone in that room and their team back at home.

The fix is speaking in the same language as your P&L. I was calling it blended-ROAS back in 2017, before the industry settled on a name. Now it has one: MER.

Marketing Efficiency Ratio; total revenue divided by total ad spend. No attribution windows, no platform self-reporting, no double-counted conversions. When Meta changes an attribution setting and your channel ROAS drops 30%, MER doesn’t move, because revenue didn’t move. It’s a bank-account balance, and that’s the point.

But revenue alone doesn’t tell you enough, either. A brand doing $10M on $2M in ad spend looks efficient until you count COGS, shipping, returns, and customer service. Contribution margin (CM2 or CM3) forces the full accounting. CM2 includes variable costs. CM3 includes marketing spend. If your CM3 is negative, your growth is subsidized, and you need to know that before your next board meeting.

CAC payback period replaces LTV as the third metric because it’s concrete. Under three months is excellent. Three to six is decent. Over twelve is a conversation with your CFO about how long the balance sheet can fund acquisition before it pays for itself.

Behind the dashboard, the validation layer is a triangle: MMM for strategic budget allocation, incrementality testing for channel validation, platform attribution for daily tactical optimization. Each cross-validates the others. MMM is already the top-rated methodology among marketers, and Google dropped its incrementality testing threshold to $5,000 in late 2025. Meta’s open-source Robyn and Google’s Meridian have made in-house modeling viable. The old objections (too slow, too expensive) are dying.

For PE-backed brands, the play is even more concrete: run a two-week holdout test on branded search. Show the operating partner that $200K of “efficient” spend generated zero dollars in new revenue. Reallocate to prospecting that drives higher top-line growth on a shorter payback period. That’s a conversation a board can follow.

Come Monday, don’t wait

The next time someone presents a ROAS number, ask one question: what happens to that metric if we pause this channel for two weeks? Four?

If nobody can answer with data, you have a reporting habit masquerading as measurement. A measurement strategy worth building is one that survives when a platform changes the rules overnight, when an AI agent bypasses your website entirely, when the conversion your dashboard attributed to paid search was going to happen anyway.

Build the toolkit. Test the assumptions. Measure whether the business is actually growing, and how.

Let me know what you find.

Leave a comment

I’m at Shoptalk this week, and the vibe has shifted.

AI has penetrated the conference floor far deeper than it had a year ago: more vendors, more integrations, more of the stack. But it’s not the breathless adoption pitch anymore. It’s a reputation exercise. “Stop the Slop.” “Hallucinations are for hippies.” The industry already bought it. Now it has to convince everyone else that was a good idea.

The most honest conversations I had this week weren’t about what AI can do. They were about what the people operating it aren’t seeing.

Here’s the number: 82% of marketers think AI benefits consumers. 42% of consumers agree. A 40-point perception gap between the people building the campaigns and the people receiving them, and 51% of marketers are now running campaigns that target AI models and web crawlers rather than humans.

The people closest to the tools have the worst read on their impact.

Enter the doom loop.

Two months ago, I wrote about the entropy problem — that messy taxonomies, mashed-together tech stacks, and product catalogs built on institutional knowledge were never designed to be reasoned with by a machine. That thesis held up this week, but the second bottleneck is worse because it’s harder to see.

We humans develop judgment by sitting with things that don’t resolve neatly. By doing the work badly, often, and usually for a long time. The eight-hundred-and-fortieth subject line teaches you something the first one can’t. The “best practices” that profitably scaled one Google Ads account were toothless in another. The campaign that tested well and bombed in market teaches you that data and judgment are not the same thing.

AI removed that friction. The AI writes the subject line now. The AI drafts the brief and the person approving the output may never have done the work that would tell them whether what they’re looking at is good, derivative, or actively damaging the brand. HBR called this the “judgment paradox”: AI simultaneously increases the need for human judgment while eliminating the messy work through which judgment is built.

The operator loses discernment, the output gets worse, rinse, and repeat. The operator can’t see what’s happening: bad judgment approving bad output generated from bad data and a bad prompt. The doom loop spirals and spirals and no one is looking up.

Anthropologie’s Anu Narayanan put it plainly from the Shoptalk stage: “Data doesn’t tell you everything. It won’t tell you what’s next.” Pinterest’s Matt Madrigal said something similar: “Taste is visual. You know it when you see it, even if you can’t describe it.” Even the platforms trying to encode taste into an algorithm admit it’s intuitive.

Judgment is not something you can shortcut, but it’s the thing we’ve been quickest to devalue.

Meanwhile, consumer enthusiasm for AI-generated content has dropped from 60% to 26% in two years. 52% of consumers actively reduce engagement when they suspect AI content. The audience is voting with their attention, and most organizations can’t hear it because only 19% track AI-specific performance metrics. They automated production without instrumenting quality.

That dissolution of trust with your consumer isn’t going to be easily identifiable in a dashboard, and by the time you do see it the damage will be difficult to undo.

The brands closing the gap.

Some brands at Shoptalk this week were telling a more deliberate story: where the human stays, and why.

Dutch Bros CEO Christine Barone took the main stage: “Emotion is the product, not the coffee.” Every brand says some version of this, but Barone is actually running the operations on it. When Dutch Bros rolled out mobile ordering, she said the time saved gets reinvested by deploying more people to engage with customers in line. Their app surfaces customer names and digital stickers on the barista’s screen, conversation starters so the human at the window is better at the thing that actually matters. They’re a sophisticated technology company (they poached Lululemon’s CTO, they have 15 million loyalty members) but the tech serves the interaction. Revenue up 28% last year, foot traffic up 13.8% while Starbucks and Dunkin’ saw declines.

Macy’s told a similar story. Their “Ask Macy’s” AI tool was built with an explicit constraint: decision-making stays with the customer, never the bot. Chief Stores Officer Barbie Cameron put it simply: “At the end of the day, it all boils down to how you make a customer feel.” They could have built an agent that recommends, selects, and checks out for you. They drew the line at recommendation, because they understand that the moment the machine decides for the customer is the moment you lose the relationship.

Then there’s New Balance, where CEO Joe Preston described 180% growth over five years and eighty new stores in 2025 — his entire growth narrative built on taste (and nostalgia), pricing discipline, and brand conviction without a single mention of AI. Sometimes the most sophisticated technology decision is knowing what doesn’t need it.

The question you should be asking.

I know the pressure. Your board wants an AI strategy. Your CFO wants cost reduction and is eyeing your headcount. The pressure to adopt fast is real and it’s coming from above. I’m not telling you to slow down, but I am telling you that the brands I saw at Shoptalk started somewhere specific: “where does judgment live in our workflow, and who owns it.”

The organizations pulling ahead right now have two things: clean data and someone in the room with enough experience to know what the tool can’t see. Judgment as an operational function, not a soft skill. The person who can look at an AI-generated campaign and tell you it’s technically sound and strategically empty, explain why, and know what to do instead.

That means asking harder questions about your AI implementation than “how fast can we scale this.” Questions like: who on your team has actually done this work by hand, and are they in the approval chain? When your AI generates a brief, who’s evaluating it? Did that person ever write one from scratch? What percentage of your AI drafts get substantially rewritten before they ship? Only 19% of companies using AI in content workflows have any measurement framework that distinguishes output quality from output volume. If your team can’t answer these questions, you’ve automated production without building the feedback loop that tells you whether it’s working.

The brands I watched win at Shoptalk this week aren’t the ones with the most integrations. They’re the ones where judgment and expertise are built into the workflow, embedded in the decisions about what to build, what to buy, what to say, what to leave out, and when.

Clean data underneath, AI enabled, and judgment woven throughout the workflow. That’s it. That’s the stack that works.

Leave a comment

In January, I wrote that the money was getting weird. Private equity firms sitting on $3.7 trillion in unsold companies, selling them to themselves through “continuation funds,” and I said it rhymed with 2008.

I was being generous.

Two months later, the United States is at war with Iran and oil hit $126 a barrel and the national debt has crossed $39 trillion and the Strait of Hormuz is functionally closed.

And that’s just the war.

There’s a lot of noise in the world right now, and I’m going to take a moment to break from my regular content to share the five things I’m seeing happening simultaneously. Not because any one of them is unprecedented, but because all five hitting at the same time is.

1. Your customer has less money and more fear

Consumer sentiment dropped to 55.5 in March. 29% of Americans now have more credit card debt than emergency savings. Macy’s posted -0.5% same-store sales while Ross Stores is up 9% and Five Below is tracking for a 12.9% increase.

The consumer is trading down. Not in the abstract “consumers are cautious” sense that gets said in every earnings call. In the specific, measurable sense that wealthier households are now shopping at Walmart — Doug McMillon said it on their earnings call — and discretionary categories are getting cut first.

If your 2026 plan assumed a stable or recovering consumer, your plan is wrong. Not pessimistic-wrong. Structurally wrong. The consumer who existed when you wrote that plan in October doesn’t exist anymore.

2. The companies you sell to, partner with, and work for are financially fragile

The PE situation I wrote about in January got worse. Fast.

Since then: Saks Fifth Avenue, Forever 21, Pat McGrath Labs, Eddie Bauer. 1,400+ store closures announced for 2026 so far. PE-backed companies now account for 71% of the largest consumer discretionary bankruptcies, up from 65% when I wrote the January piece.

Two months. Six percentage points.
The. actual. f*ck.

But here’s the part that didn’t exist in January: the $265 billion private credit meltdown. The shadow banking system that funded the PE buyout wave is seizing up. Apollo down 41% from peak. Blackstone down 46%. KKR down 48%. Blue Owl down 66%. Blackstone had to inject $400 million of its own capital into one of its credit funds to stop investors from running for the exits.

This is how the PE problem becomes your problem. They can’t exit. They can’t refinance — not with $1.35 trillion in corporate debt maturing this year. So they squeeze the companies they own. Marketing budgets get chopped. Agency retainers get canceled. Headcount freezes, then layoffs.

And then there are the zombie brands. PE-backed companies that are essentially insolvent but still spending aggressively on customer acquisition, trying to show growth before a restructuring that everyone in the room knows is coming. You’re competing for the same customer against a brand that is lighting investor money on fire. That’s not a competitive landscape. That’s a burning building with a sale sign in the window.

3. Your team is getting smaller. The people being laid off are your customers.

108,435 layoffs were announced in January 2026 alone. The highest January total since 2009. Hiring plans are down 63% year over year.

Meta is cutting 20% company-wide. Fifteen thousand people out the door while committing $135 billion to AI infrastructure. Amazon cut 16,000 in January. UPS is replacing 30,000 jobs with automation. DOGE has eliminated 242,000 federal positions, ten percent of the civilian workforce. For the year to date, 12,304 layoffs have been directly attributed to AI replacement.

The professional-managerial class being laid off is the exact demographic that premium B2C and B2B brands target. And the people still employed? They’re not switching jobs, not taking risks, not spending with confidence.

4. The market is pretending everything is fine

Ad spend forecasts (the ones published before the war) were actually up. Madison and Wall boosted their 2026 US forecast to 10.2% growth. Magna projected global ad revenue hitting $979 billion.

Those forecasts were written in a different world than you and I are living in, dear reader.

The reality: 42% of marketers now anticipate lower budgets in 2026, up from 22% last year. WARC cut $19.8 billion from global ad spend forecasts. Three months before the 2003 Iraq War, 98% of advertisers felt confident they’d hit their sales targets. By the time the war started, 82.5% expected to miss.

That pattern is repeating. The money isn’t disappearing evenly. It’s consolidating into channels with defensible ROI and draining from everything else. If your channel can’t prove its math to a nervous CFO in a single slide, your channel is at risk.

5. Remember what Anthony Bourdain taught you about CDOs?

In 2008, the transmission mechanism that turned a housing problem into a global financial crisis was the CDO — opaque, illiquid, supposedly diversified instruments that turned out to be concentrated in the same correlated risk. Nobody knew what was in them until it was too late. The rhyme scheme is getting uncomfortably precise.

In 2026, the equivalent is private credit. A $1.8 to $2.1 trillion market with the same features: opacity, illiquidity, and extreme sector concentration. Default rates have climbed to 5.8%, the highest level ever tracked. And they’re building Private Credit CLOs on top of this. Leverage on leverage.

The war didn’t create this fragility — it detonated it. The Gulf sovereign wealth funds that were the most reliable funding source for private credit now need emergency liquidity at home. The Strait of Hormuz closure cratered their oil export revenue. S&P Global flagged $307 billion in potential deposit flight from GCC banks. They can’t sell their illiquid positions without fire-sale losses. Trapped on both sides.

And here’s the part that connects this to your customer’s kitchen table: Blackstone, Apollo, and Ares have been pushing private markets into 401(k) portfolios — targeting the $13 trillion in American retirement accounts. If retail investors start to perceive that their retirement savings are trapped in the same gated, illiquid structures, the resulting panic won’t stay in the financial pages.

This is where I’ll leave you, for now

None of these things, in isolation, would be unprecedented. We’ve had wars. We’ve had debt crises. We’ve had layoff waves. We’ve had tariff regimes.

We haven’t had all of them at the same time, hitting the same consumer, the same supply chain, and the same budget simultaneously. Each one makes the others worse. The war drives oil prices up, which drives consumer costs up, which drives sentiment down, which drives spending down, which drives budgets down, which drives layoffs, which drives spending down further. It’s a loop. And it’s tightening.

This isn’t just a recession. The last recession was a cycle — things contract, things recover, the line goes back up. This is more a structural shift. The system that’s been propped up since 2008 with cheap money, cheap debt, cheap labor, and cheap attention is running out of props. And instead of fixing the foundation, the people with the most to lose broke the last load-bearing wall and blamed it on the machines.

I’m a millennial. This is my fourth major economic crisis before 40. Every time my generation starts to get our footing, the floor drops again. At a certain point you stop calling it bad luck and start calling it what it is: a system that was never built for us to win.

Putting that aside, though, the best case scenario right now, the most optimistic credible forecast I can find, is a “brief growth recession” followed by a rebound. That’s the good version. That’s the version where everything goes right.

No ad budget is going to fix what’s happening right now. No marketing strategy is going to paper over $126 oil and 108,000 layoffs and a tightening credit market. I’m not going to insult you with a five-point action plan.

What I’m doing — with my own hands, my own money, my own time — is building. Tools. Infrastructure. Things I own and control, because the systems we’ve been renting from platforms and vendors are about to get repriced, gated, or killed.

I’m learning things that scare me. Sitting with code I don’t fully understand, breaking things, building them back wrong, because the gap between what I can do now and what I’ll need to do in 8 months is not closeable by watching webinars.

I’m watching the same playbook run everywhere: companies cutting the teams that build and keeping the teams that report, optimizing for a consumer who no longer exists, and holding still and hoping the ground stops moving.

Spoiler: It doesn’t stop.

Every crisis leaves a crater. The 2008 crash left one in consumer trust, and those who filled it (Warby Parker, Shopify, Stripe) didn’t wait for the old market to come back. They built in the wreckage while everyone else was refreshing their 401(k) balance.

I don’t know what the crater looks like this time. But I’d rather be the one with a shovel than the one still arguing about why the ground moved.

Leave a comment

She’s not a chatbot, nor is she OpenClaw. She’s 60 TypeScript files, 9,500 lines of code, 40 tools, and a security architecture I spent days hardening. She monitors my inbox, drafts emails, watches my Docker containers, checks my calendar, makes appointments. I can text her on Telegram from my phone and she’ll spin up a Claude Code session on the Mac Mini and send me back a direct link for remote control so I can skirt API costs.

She’s genuinely useful.

She was the easy part.

The hard part was the moment she needed to coordinate with other systems. The moment her output became someone else’s input. The moment I went from “I built an agent” to “I built a system.”

That’s the orchestration layer, and I’m seeing so few talking about it beyond niche circles.

The Single-Agent Illusion

Here’s what every agent tutorial looks like: you give an LLM some tools, maybe a system prompt with a persona, and you let it run. It’s powerful. It’s fun. It works.

Until it doesn’t.

The first time I needed multiple agents to collaborate on the same codebase — one reading a roadmap, one decomposing tasks, one writing code, one reviewing it, one managing the merge queue — I realized I wasn’t building an agent team anymore. I was building a system, and systems have entirely different failure modes.

The demo version of AI is one agent, one task, one session. The production version is five agents, three handoffs, two quality gates, and a daemon running 24/7 with no human in the loop.

Those are not the same job.

UC Berkeley published a paper last week analyzing 1,600+ failure traces across seven multi-agent frameworks. Their finding: 79% of failures were coordination and specification problems — not model capability, not infrastructure, not prompting.

They built a whole taxonomy for this — the Multi-Agent Systems Failure Taxonomy (MAST) — and the breakdown is damning. Upwards of 40% of failures were specification issues. Another ~40% were inter-agent misalignment. The specific failure modes read like a postmortem of every bug I’ve shipped in the last week:

• Conversation Reset (FM-2.1): An agent loses its context mid-task and starts over, undoing its own progress. I’ve watched this happen in real time — an agent 80% through a code review just... forgets what it was reviewing.

• Task Derailment (FM-2.3): The agent drifts from the actual objective into circular, irrelevant work. This is the one where your dev agent decides to “refactor for consistency” instead of implementing the feature you actually asked for.

• Clarification Failure (FM-2.2): The agent proceeds with wrong assumptions instead of asking for help — responsible for another ~12% of all system failures. This one’s insidious because the agent looks productive the whole time. It’s confidently building the wrong thing.

These aren’t “bugs” in the LLM. They’re emergent properties of the orchestration layer.

The agents were smart enough. The systems weren’t.

What Actually Breaks (and How to Fix It)

I’ll save you the suspense. It’s not the AI. The LLM calls work fine. What breaks is everything around them — the boring, un-demoable infrastructure that nobody puts in their Twitter thread.

Here are the four orchestration problems I’ve hit repeatedly, and the patterns I’ve landed on for each.

1. Auth and Credentials

The problem: Claude’s CLI needs macOS keychain access for OAuth. Sounds trivial. But my agents run as background daemons via LaunchAgent, and SSH sessions can’t access the keychain. I spent a full day debugging why an agent task — which worked perfectly when I ran it manually — silently failed at 10 AM on a Sunday.

This got worse with macOS Tahoe. The security command-line utility — the standard way to read keychain secrets — now frequently hangs indefinitely or returns exit code 36 when called from a LaunchAgent. Background processes don’t get a SecurityAgent session, and Tahoe’s tightened ACLs mean a secret stored by one binary path may be invisible to another. If you’re running agents as daemons on a Mac in 2026, this will bite you. It bit me.

The fix: launchctl kickstart instead of SSH. Every claude --print invocation has to go through the LaunchAgent so it inherits the GUI session’s keychain. I hit this bug weekly until I documented it in all caps: MUST use launchctl kickstart (not SSH nohup) — Claude CLI needs macOS keychain for OAuth.

The pattern: Any time an agent invokes another service that requires credentials, the execution context matters as much as the credentials themselves. Test your agents in the exact environment they’ll run in — not your terminal, not your SSH session. The daemon context.

This isn’t an AI problem. It’s a systems problem that just happens to have AI in the stack.

Those of us without engineering backgrounds are starting from behind the line, and all of this may sound like Latin. It sounded totally foreign to me a few months ago, too.

2. Error Handling: Stop or Continue?

The problem: My orchestrator chains multiple agents together: a PM agent reads the roadmap and submits goals, a planner agent decomposes them into tasks, a dev agent writes the code, a QA agent reviews it. If the planner fails — bad input, parsing error, whatever — should the dev agent still pick up the last set of tasks?

In a single agent, the answer is obvious: crash and tell me about it. In an orchestrated chain, it depends. The dev agent can work on existing tasks even if no new ones get planned. Stale tasks are better than idle agents, but a corrupted task spec is worse than nothing — the dev agent will write code to the wrong spec and the QA agent won’t catch it.

The fix: Each handoff gets its own failure policy. The planner failing is a warning — log it, keep going, existing tasks are still valid. But if the dev agent produces corrupt output (syntax errors, failed tests, etc), hard stop. Do not pass go. Do not pass garbage to QA.

The pattern: Every handoff in a chain needs its own failure policy. Ask three questions at each link:

1. Can the next step run with stale/missing input?

2. Is stale output better than no output?

3. Will a failure here corrupt downstream work? My system has five links:
   
   trigger → planner → dev → QA → merge

and each one answers those questions differently.

3. State Management

The problem: Every task in my system has a status:

backlog → claimed → running → review → done

The agent loop picks up backlog tasks, claims them, executes, and updates the status. Simple enough.

But what happens when an agent crashes between claiming a task and committing the code? The task stays claimed. Tomorrow’s loop sees it and skips it — it’s claimed, someone’s already working on it. But nobody actually is. Now you’ve got a stuck task and an agent that thinks it’s idle.

Or worse: two agents grab the same task in the same polling cycle. Now you’ve got conflicting branches, duplicate work, and a merge queue that can’t resolve the conflict.

The fix: SELECT FOR UPDATE SKIP LOCKED. Agents claim tasks atomically at the database level — if another agent already has it, the query returns nothing instead of blocking. And stuck tasks get a timeout: if a task stays claimed for longer than its trust tier’s max turns would allow (more on that in a moment), the system releases it back to the queue.

The pattern: State transitions should be atomic — claim the task and update the status in a single database transaction, so a crash produces a recoverable state rather than a phantom claim. This is solved-problem territory in software engineering — but the agent ecosystem hasn’t caught up yet. Most agent frameworks treat state as an afterthought, and honestly, most tutorials pretend it doesn’t exist.

I’m not the only one hitting this wall. A developer named Patrick runs a 5-agent system 24/7 on a Mac Mini — eerily similar hardware to mine — and independently landed on the same architecture: three files (current-task.json, context-snapshot.json, outbox.json) as the source of truth for every agent in the system. No framework. Just files and discipline. His conclusion mirrors mine: 80% of production agent failures trace to state management, not prompt quality.

4. Quality Gates

The problem: Here’s the loop my orchestrator runs for every dev task:

1. A dev agent writes code in an isolated git worktree

2. A QA agent reviews the code against the task spec

3. If QA says “revise,” the dev agent gets the code back with inline comments

4. Repeat up to 3 times

5. If it passes, it enters the merge queue

This is not one agent. This is three separate LLM calls with different system prompts, different evaluation criteria, and a control loop managing the transitions. The dev agent doesn’t know about the QA agent. The QA agent doesn’t remember the previous review. State flows through branches and task records, not through the model’s context.

The fix: The quality gate is the most important part of the system, but it’s the hardest to get right. The first time I let an agent write code with no QA gate, it reformatted my entire project structure “for consistency.” Another time, an agent passed its own QA review and merged code that broke the build — the QA criteria were checking for style and correctness but not for integration with the rest of the codebase.

I had to add trust tiers. New agents start as interns:

intern → junior → senior → principle

They earn promotion to junior, senior, and principal through consecutive successful completions — not set configuration. Three consecutive failures trigger a demotion. Trust isn’t a setting, it’s not unconditional. The agents have to earn it and prove they can keep it.

The pattern: Your quality gate is only as good as its criteria — and your trust model is only as good as its consequences. Build gates that check what actually matters (does this code integrate? does it pass tests?), not just what’s easy to check (does it look clean?). And give agents less access until they’ve proven they won’t trash the repo.

But trust tiers alone aren’t enough — and this is something I’m still building toward. The emerging standard in production AI governance is decision lineage (what I was calling the ‘judgement layer’ last week): every autonomous action generates a log documenting exactly why the decision was made, which data was accessed, and which policy was applied. Not just “agent committed code” but “agent committed code because task #47 specified this change, QA passed on attempt 2 of 3, and the agent’s trust tier authorized write access to this repo.”

The bigger idea here is Non-Human Identity (NHI) — giving every agent its own audited identity separate from your credentials. Right now, when my principal agent pushes code, it pushes as me. Even a principal-tier agent has to pass through a pre-push security gate that scans for secrets and PII before anything leaves the machine — but it’s still my name on the commit. That’s fine for a solo project. It’s a liability for anything regulated. The industry term for the alternative is “shadow autonomy” — agents acting under a human’s credentials with no way to distinguish who (or what) actually made the call. Decision lineage kills that ambiguity.

From Agent to Orchestrator

This is where claw.clip1 comes in.

After a couple weeks of building and using Aouda — and watching the orchestration question get bigger with every new capability — I started building a dedicated orchestrator. Not a framework. Not a library. A daemon that runs 24/7 on the same Mac Mini, managing teams of Claude Code agents.

claw.clip assigns tasks from a priority queue. Agents claim work atomically — remember: SELECT FOR UPDATE SKIP LOCKED — so two agents never grab the same task. Trust tiers are invoked.

A PM agent runs every 30 minutes. It reads the roadmap — an actual ROADMAP.md file in the repo — and submits incomplete items to a planner agent, which decomposes them into 10-30 minute tasks. The PM doesn’t write code. It writes task specs. Other agents pick them up.

And here’s the part that still feels a little surreal:
claw.clip is building itself.

The agents working on improving the orchestrator are being orchestrated by the orchestrator. It tracks their costs, enforces their budgets, manages their merge queue, and promotes them when they deliver clean work.

Is it overkill for a personal project? Potentially.

But it taught me something I couldn’t have learned from a tutorial: orchestration isn’t a feature you add to an agent. It’s a different discipline entirely.

Where This Goes Next: A2A and MCP

Right now, my orchestration is local. Agents talk through files, database rows, and git branches — all on one machine. That works for a solo operator. It doesn’t work when your agents need to collaborate with agents built by someone else.

The industry is converging on two protocols that solve this: Agent2Agent (A2A) and Model Context Protocol (MCP).

A2A is the messaging layer — it defines how agents find each other and hand off work. Every agent publishes an “Agent Card” (a machine-readable JSON file advertising its skills and auth requirements), and tasks move through a standardized lifecycle:

Submitted → Working → Completed or Failed

If that sounds familiar, it’s because it’s basically what my task queue already does, except it’s a universal standard instead of my personal PostgreSQL schema.

MCP is the tool layer — it standardizes how agents access tools and data sources. Instead of hardcoding every integration, an agent discovers available tools at runtime through MCP servers. I’m already using MCP for Aouda’s tool access. The gap is that my task output format doesn’t yet align with A2A’s Artifact schema — which means my agents can talk to each other, but they can’t talk to yours (yet).

That’s the next frontier. Local orchestration is table stakes. Interoperable orchestration is where this gets interesting.

The Career Reframe

Every tutorial, every demo, every “I built an AI agent in 20 minutes” post is about the single agent. Build it, ship it, done.

But production doesn’t work like that. Production is: what happens when this agent needs to hand off to that agent? What if the second one fails? What if the third one produces garbage and the fourth one can’t tell? What if the whole thing fails at 6:30 AM while nobody’s watching and all agents sit idle?

Three years ago the hot-topic hiring filter was “can you prompt an AI.” Three weeks ago it was ”can you design an agent”. I’m going to go ahead and say that a month from now, everyone will be talking about how you need to know how to build an orchestration layer.

Just watch.

It’s not a hypothetical, it’s starting to happen: “Agent orchestration specialist” is now a named job category. Eightfold AI called it the most important job of 2026. BMW is hiring a “Lead Agentic AI Orchestration” role. Job postings requiring agentic AI skills grew 986% between 2023 and 2024.

Prompt engineering is one thing, but systems engineering is entirely different animal. It’s state machines and error policies and trust hierarchies and quality gates. It’s set -uo pipefail vs set -euo pipefail. It’s knowing that SSH can’t access the keychain no matter how many times you forget to run LaunchAgent.

It’s the boring stuff, and it’s where the actual value lives. The good news is that this is learnable.

Leave a comment

1

The name is a nod to Paperclip, an open-source project that beat me to market with the same idea — orchestrating teams of AI agents with org structure, budgets, and governance. Great minds, etc. Where Paperclip is a cloud platform designed for enterprise teams running agents from any provider, claw.clip is a daemon — it runs 24/7 on bare metal, spawns Claude Code sessions directly via the CLI, and manages everything through files and git worktrees. It’s built for a solo operator on her own hardware, not a team on a cloud dashboard. I have much to do before it’s ready for a public repo, but follow along if you want to check it out as soon as it’s live.

---

D2C Is Dead. Welcome to A2A.

Every letter in this naming convention is a human leaving the room.

D2C cut out the retailer. B2C widened the tent. B2B served the enterprise. B2A — the shift I wrote about in January — meant your customer sent an AI agent instead of showing up themselves.

Now: A2A. Agent-to-agent commerce. Your brand’s AI negotiating with your customer’s AI. No human on either end — not in the browsing, not in the negotiation, not in the checkout.

The sale closes and neither party is a person. Sit with that.

Last week I wrote about what brand looks like when it becomes a spreadsheet. This is final essay in the Agentic Shift series is about what happens on the other side of the table — when the buyer isn’t there either.

“Agents negotiating with agents” sounds abstract until you walk through it, and it’s something I find both awe-inspiring and terrifying.

What A2A Actually Looks Like

Someone — let’s call her Maya — wants to find a birthday gift for her friend. She opens her phone and tells her personal assistant: “Find me something good under $50. She likes ceramics and sustainable brands.” Maya doesn’t open a browser. She doesn’t open Instagram or Pinterest. She doesn’t visit your website. She doesn’t see your ad. She gives intent to her agent and goes back to her coffee. (I picture her at a kitchen counter, still in yesterday’s socks, not thinking about your brand at all.)

Her agent knows her preferences, her purchase history, her friend’s publicly available wishlist. It uses an MCP — the protocol that standardizes how agents connect to external data — to query merchant catalogs. Not typing keywords into a search bar. Issuing a structured request: category, price ceiling, brand criteria, available inventory. Dozens of catalogs, queried simultaneously, in the time it takes Maya to pour a second cup.

The reason this works now and didn’t three years ago is a protocol problem that finally got solved. Before MCP, every agent needed a custom integration for every data source — the complexity scaled as N-times-M and made the whole thing slow, expensive, and impractical. MCP collapsed it to M+N. Web MCP means doesn’t even need to take screenshots to read a page anymore. One integration, every agent ecosystem. That’s the unlock.

Your brand’s agent — the one Klaviyo calls a Customer Agent — receives Maya’s request. It doesn’t return a generic product page. It pulls Maya’s friend’s profile data from your SSOT (if she’s a past customer), surfaces the ceramics that fit her aesthetic, and packages an offer: a $44 ceramic planter, gift-wrapped, with expedited shipping. Maybe a 10% first-purchase discount because the query came from a new household. Maybe a bundle. It negotiates.

Maya’s agent evaluates the response against other offers it received in parallel. It factors in reviews, return policy, delivery window, brand trust signals. It ranks. It selects.

Then the plumbing kicks in. ACP — the Agentic Commerce Protocol, developed by OpenAI and Stripe as an open standard — manages the cart and checkout sequence.

The money moves through AP2, Google’s Agent Payments Protocol, built with PayPal, Mastercard, and 60+ partners. AP2’s key innovation is the mandate: a cryptographically signed digital contract proving Maya authorized her agent to make this purchase on her behalf. Tamper-proof. Verifiable. No fraud ambiguity. No disputed charge. No human hand on the keyboard at any point.

The planter ships. Maya gets a notification. Her friend gets a gift she actually wanted.

Your brand made a sale without a single impression, click, or email open.

What “Selling” Becomes

Here’s where it gets uncomfortable for people who built careers on persuasion. (People like me.)

The whole infrastructure of marketing was built on the assumption that there was a person at the end of it who needed to be *moved*. Copywriters wrote headlines because humans respond to language. Art directors chose images because humans respond to visual cues. CRM teams built journeys because humans respond to timing and relevance.

In A2A, there isn’t a person at the end. Maya’s agent isn’t moved by your brand story. It doesn’t feel the pull of aspirational lifestyle photography. It doesn’t respond to urgency copy (”Only 3 left!”). It evaluates:

• Does this product match the stated parameters?

• Is the brand’s data clean and structured?

• Does the trust signal — reviews, return policy, fulfillment reliability — meet threshold?

It’s not persuasion. It’s evaluation. And the difference changes everything about where you put your effort.

The creative brief doesn’t die. It becomes the agent’s personality — the training data that determines how your brand agent answers questions, what it prioritizes, how it negotiates. Here’s what that actually looks like: where a creative brief used to say “Our voice is warm, knowledgeable, and slightly irreverent — think the friend who knows wine but won’t be weird about it,” a model specification says “When price queries fall within 15% of ceiling, lead with product-match confidence score, then social-proof signal, then offer a curated bundle. Tone: consultative. Never open with discount.” The brief stops being a document for a copywriter and becomes a specification for a model. Same brand DNA, different container.

Customer relationships become agent relationships. Loyalty isn’t emotional anymore — it’s contextual. Does your agent remember that this customer prefers matte finishes and hates bubble wrap? (The bubble wrap thing is real. I have opinions about unnecessary packaging and I want my future AI to know them.) Does it know she shops for her mother-in-law every December and her price ceiling goes up 30% for those orders? Does it negotiate well, or does it return generic offers that Maya’s agent deprioritizes immediately?

Clean data and a well-trained agent aren’t nice-to-haves.
They are the product.

The marketer’s job shifts from “create campaigns that persuade humans” to “train agents that negotiate competently on behalf of the brand.” That’s closer to product management than copywriting. It’s closer to data architecture than brand strategy. It requires caring deeply about things that used to live in the engineering org: schema consistency, API reliability, protocol compliance. I think the people who are slightly annoyed right now — the ones thinking *I didn’t get into marketing to care about schema* — are the ones who most need to hear this.

This is the part where I’m supposed to reassure you that creativity still matters. It does. But the creativity that matters in A2A is the creativity of structuring information so that an agent can use it — not the creativity of making a human feel something. Those are genuinely different disciplines. The people who understand both will be worth a lot.

The Through-Line

D2C was about getting closer to the customer. We spent a decade celebrating that proximity — the customer relationship as competitive moat, the first-party data as the thing that made you defensible.

A2A is the inverse. The customer isn’t closer.
The customer isn’t even there.

What’s there instead is a system that acts on her behalf, makes decisions according to her stated and inferred preferences, and executes without her involvement. The relationship isn’t gone — her preferences, her history, her trust in your brand all still feed into the outcome. But the moment of transaction has been delegated entirely.

For people who built careers on the belief that great marketing is fundamentally about human connection — this is a genuinely disorienting thing to sit with. I don’t think it means human connection stops mattering. I think it means the place where that connection gets established moves earlier in the funnel, further from the transaction, deeper into brand infrastructure. Changes how brands show up and add value away from the purchase cycle.

You earn the relationship before the agent ever shows up.

The agent executes it.

Which means the question isn’t whether you believe in A2A. The question is whether your brand data is clean enough, your agent is trained well enough, and your infrastructure is structured enough for someone else’s AI to choose you when no human is watching.

That’s a harder question than “is our creative good?” And most marketing teams aren’t asking it yet. I’d bet dollars to donuts they will be by next quarter.

p.s. There’s a lot of snake oil salesmen in the AI space right now. I’ve been working on a few things and will be putting my money where my mouth is with a few releases coming to an open source repo near you. Stay tuned for more

Leave a comment

.

The person running your CRM is probably very good at their job. That’s not the problem. The problem is that the job is becoming something else — and the skills that made them excellent at the old version don’t automatically transfer to the new one.

They were hired to manage email flows and segmentation. A/B test subject lines. Build a welcome series. Suppress unengaged contacts before a big send.

I’d wager that within 18 months, that same platform will be the layer where your brand’s AI agent and your customer’s AI agent negotiate recommendations, manage cart logic, and close transactions without a human touching the keyboard on either side.

Those are not the same job.

This isn’t a threat piece. It’s an opportunity map. The gap is real. It’s large. And it’s closing faster than anyone is preparing for. The marketer who closes it first isn’t in trouble — they’re in an extremely good position.

What Your CRM Is Actually Becoming

Your customer doesn’t want to talk to you anymore — not to your chatbot, not to your support queue, not to your drip campaign. They want to send their agent to deal with yours. I wrote about that shift in January. This piece is about what it means for the infrastructure sitting at the center of your marketing operation.

We’ve been calling it a CRM — customer relationship management — for so long that the name has become a liability. It implies a tool for managing relationships between humans. A place to store contacts. A mechanism for sending scheduled messages to segmented lists.

That’s what it was. It is not what it’s becoming.

The shift underway is from CRM-as-messaging-tool to CRM-as-SSOT: Single Source of Truth. Not a platform that humans query through a dashboard to build campaigns, but a structured data layer that AI agents query in real time to negotiate, transact, and personalize at machine speed. IDC projects that by 2026, nearly half of new CRM investment will go to data architecture and AI infrastructure rather than additional licenses or modules. The money is following the architecture.

The primary user of your CRM used to be a marketing specialist scheduling sends. It is becoming an autonomous AI agent pulling live signals to make decisions that used to take a human three clicks and a gut check.

Relational databases optimized for dashboard readability are giving way to structured endpoints designed for machine access. Scheduled blasts are giving way to real-time exchanges between brand-side and consumer-side agents — adjusting offers, handling objections, closing transactions in the time it used to take your automation to fire a trigger email. And the success metric that matters is no longer open rate or click rate.

The metric you need to be measuring is inference advantage — how much better your agent performs because it has richer, more current, more accurate data than the agent on the other side. That’s not a metric most lifecycle marketers have ever been asked to think about, let alone optimize.

Two platforms are worth naming because the specifics matter.

Klaviyo moved first and moved fast. In August 2025, they launched an MCP server — a standardization layer that allows third-party AI agents to connect directly to your Klaviyo instance. Claude, ChatGPT, whatever agent stack you’re running — they can now query your performance metrics, segments, and campaign history using natural language. No custom integration. No API connector built by your dev team. One standardized interface. They also shipped a Customer Agent inside K:Service — a brand-side AI that handles customer interactions across chat, SMS, and email, with full access to your first-party data and personalization logic.

Salesforce is building toward the same model with Agentforce, though they’ve had a rougher ride. In September 2025, Noma Security disclosed ForcedLeak (CVSS 9.4) — a prompt injection vulnerability that let attackers embed malicious instructions in Web-to-Lead forms, causing AI agents to exfiltrate sensitive CRM data. The security architecture for agentic interaction is still being built in real time.

The underlying protocol structure is what changes the scale equation. Before standards like MCP existed, every agent needed a custom connection to every data source — the classic M×N integration problem. MCP reduces that to M+N: implement once on each side, be accessible to every major agent ecosystem. Layer in the Agentic Commerce Protocol (ACP, maintained by OpenAI and Stripe) and Google’s AP2 for payment authorization, and you’re looking at a fundamentally different architecture for how brand data moves through the world.

The Skills Gap

Here’s what I keep thinking about.

The lifecycle marketer who has spent five years getting very good at segmentation strategy, flow logic, subject line testing, send-time optimization, and suppression management has built a genuinely valuable skill set. None of that was wasted. But that skill set was built for a platform that delivered messages to human inboxes and measured whether humans opened them.

The skill set that the same platform now requires looks different:

Protocol negotiation* Understanding how MCP endpoints are structured, what agents can and can’t access, and why that matters for what your brand can offer at the moment of intent.

Agent authentication* understanding how your brand-side agent proves it has authorization to act, and what happens when it doesn’t.

Inference optimization* structuring your first-party data so your agent has a material advantage in a negotiation with a consumer’s agent that has its own data signals.

Data governance as a real-time problem* not a quarterly hygiene exercise.

These are not impossible skills to develop. (They’re not even that exotic — a marketer with strong technical instincts and intellectual curiosity can get there.) But they are not what your current team was hired for, and they are not what most lifecycle marketers are building toward right now.

The organizations that close this gap first won’t be the ones that replace their marketing teams with engineers. They’ll be the ones that identify which marketers have the instinct for this kind of technical fluency, give them room to develop it, and build roles that didn’t exist two years ago.

The Security Layer Nobody Has Staffed For

The ForcedLeak vulnerability exposed something the whole industry will have to reckon with: agent-to-agent communication creates attack surfaces that traditional marketing security models weren’t built to handle. When your brand’s agent is processing data through a shared protocol layer, the question of what gets exposed to whom is not a simple one.

This used to be a security team problem. It is becoming a lifecycle marketing problem — because the people who know your customer data most intimately, how it’s structured, what’s used for personalization, what’s sensitive — sit on the marketing side. And nobody has asked them yet which of that data should be queryable by an external agent. (I’m going deeper on this in the next piece. For now: if nobody on the marketing side has been in a room with security to discuss protocol-level data exposure, that conversation is very long overdue.)

The Career Opportunity

Gaps are where careers are made. Not gradually, through steady advancement — but in the moment when everyone else is still reading the old map.

The shift from CRM-as-email-tool to CRM-as-agentic-SSOT is creating that gap right now. The platforms are ahead of the practitioners. The infrastructure is moving faster than the job descriptions. The roles that will exist in 2028 — growth engineer, agent orchestration strategist, agentic data architect, lifecycle protocol manager, whatever we end up calling them — are not common yet. But the people who will fill them are already working somewhere, probably debugging why a segment isn’t behaving the way it should.

The marketers who get there first won’t be the ones who waited for a job description to tell them what to learn. They’ll be the ones who looked at what their platform was becoming and started asking: what would I need to understand to be useful in that world?

Protocol architecture. Agent authentication. First-party data governance. Inference advantage. These aren’t developer skills — they’re strategic skills, and they sit exactly at the intersection of marketing judgment and technical fluency that good lifecycle marketers already operate in.

The job you were hired for may not exist in its current form for much longer. The platform does. The question is whether you’re building toward what it’s becoming.

*Up next in the Agentic Shift series: The transaction — what commerce looks like when neither party is human, and why the marketer’s role shifts from campaign creator to agent trainer.

An AI agent doesn’t care about your origin story.

Here is the thing no one in a brand strategy deck will say out loud: the agent doesn’t care that your supplement brand was founded by a triathlete who couldn’t find a clean protein powder. It doesn’t care that your headphones are inspired by “the way music should feel.” It reads your structured data, compares it against eleven other options, matches it to a user query, and either includes you or it doesn’t. The decision takes milliseconds. Your brand narrative — the one you paid a consultant $40,000 to develop — is not a variable in that equation.

I wrote in January about what happens when your customer stops wanting to talk to you. In Your Customer Doesn’t Want to Talk to You Anymore I asked,“What does brand even mean in that interaction?”

This is the answer.

The Reduction

When an agentic system — OpenAI’s shopping agent, Google’s AI Mode, a custom GPT wired to a procurement workflow — evaluates your product, it is doing something that feels reductive because it is reductive. On purpose. That is the function.

It compares structured attributes: price, shipping window, verified ingredient purity, return policy terms, carbon footprint certification, third-party review aggregates, etc etc. It’s working from a schema, not a vibe. And the schema doesn’t have a field for “brand voice.”

What it does have fields for: everything you either filled out completely or left blank.

This is where it gets specific. If a user asks an agent for “wireless headphones with active noise cancellation under $200” and your product data doesn’t explicitly specify the *type* of noise cancellation — hybrid ANC, feedforward, feedback — in a format the agent can parse, you get filtered out. Not ranked lower. Filtered out. The agent doesn’t penalize you for being vague. It simply doesn’t see you.

That distinction matters. Because legibility is a skill. It’s buildable. And most brands aren’t building it.

But legibility without genuine product quality is a dead end — structured data can get you into the consideration set, but it can’t survive a bad review aggregate or a return rate that tells its own story. The machine reads everything, including the parts you wish it wouldn’t. This isn’t a hack. It’s a discipline.

Attribute Completeness Is the New Brand

A mid-market pet supplement company — real products, decent reviews, a founder story involving a sick rescue dog — implemented a comprehensive Answer Engine Optimization strategy. High-density schema markup. Third-party entity validation across every data layer an agent might query. Complete attribute specifications at the SKU level. Within four months: 269% increase in LLM-referred traffic and an 185% increase in organic traffic.

They didn’t redesign their logo. They didn’t hire a new agency. They made themselves legible.

An Every Day Carry brand did the same — optimized specifically for agentic query resolution, structured their product data for AI inference rather than human browsing. AEO agencies are now reporting average traffic increases of 920% from AI-driven channels across their client portfolios. The EDC brand’s results tracked with those numbers — eleven times the purchases.

Eleven times.

These are not edge cases. These are early movers in a shift that’s already here. These number will only continue to grow with AI adoption.

In February 2024, Gartner predicted traditional search engine volume would drop 25% by 2026 due to AI chatbots and virtual agents — with organic search traffic declining 50% or more by 2028. We’re now inside that window. Zero-click searches account for over 60% of queries, and when AI Overviews trigger, that number climbs past 80%. The query gets answered in the interface. Your product either shows up in that answer or it doesn’t. And showing up is a function of how completely your data is represented — not how beautiful your brand guidelines are.

The framing I keep coming back to — and the one I’ll use throughout this series — is *inference advantage*. Not “win the click.” Win the inference. Be the source an AI cites, not a result it lists. Inference advantage means your structured data is complete enough, validated enough, and contextually rich enough that the agent selects you during its reasoning process. That’s the new competitive surface.

Brand used to mean the story you told. Now brand also means how completely your attributes are represented in every data layer an agent might query. Those two definitions aren’t in conflict. But the second one is no longer optional.

The Bilingual Problem

Here is where it gets genuinely interesting — and where most marketing teams are going to have to split.

You still have human customers. As of now, most of your customers are still human-initiated — browsing, scrolling, reading reviews, clicking around at 11pm because they can’t sleep and they need a new water bottle. Those customers respond to story. They respond to aesthetic. They respond to the particular quality of light in your product photography and that influencer they like who mentioned you and whether your copy sounds like a person or a press release.

The best marketers are going to be bilingual. Fluent in human emotional resonance and machine readability simultaneously. Not choosing between them. Running both systems in parallel, understanding where each one applies, and building infrastructure that serves both without either cannibalizing the other.

That’s a different skill set than what most marketing job descriptions ask for. If I may be so bold, it’s also a more interesting one.

There’s a counter-movement worth acknowledging — what some people are calling the Acoustic Movement, though it doesn’t have a stable name yet. A subset of premium and luxury brands are going in the opposite direction entirely: deliberately making their digital presence AI-inaccessible. Gated human verification. Non-standard typography that’s beautiful and intentionally unparseable. Physical-only catalogs. Prada staged its Fall/Winter 2025 collection as what Miuccia Prada described as “a reaction to the first season of artificial intelligence.” Tod’s debuted an “Artisanal Intelligence” campaign spotlighting human hands over algorithms. Liquid Death’s Olympics ad that leaned into AI skepticism and was actually pretty scary.

High friction as a status signal (kind of love it, tbh).

The argument is that AI accessibility is mass-market, and mass-market is death for luxury positioning. Being machine-hostile becomes proof of exclusivity. I think this is a coherent strategy for a specific tier of brand. Hermès can pull it off. Your DTC skincare line probably cannot. And even if it works as positioning, it doesn’t eliminate the underlying shift — it just opts out of one distribution channel and bets that the human channel stays premium indefinitely. That’s a bet I wouldn’t make at scale. But it’s a real option, and intellectually it’s a fascinating one — the brand that proves its value by being unreachable to the machine.

Most brands don’t get to make that bet. Most brands need the agent to find them. Which means the work is bilingual fluency, not retreat.

This Is Not a Crisis

Here is the reframe I want to leave you with.

Every few years, some new layer of infrastructure rewrites what brand-building requires. The web arrived and suddenly brand required a website — not just packaging. Social arrived and suddenly brand required a content calendar and a community manager. Search arrived and suddenly brand required SEO — which, at the time, felt to a lot of creative marketers like a betrayal of the craft. “You want me to write for a robot?”

Fun fact: I was that creative marketer.

But we wrote for the robot. And then we wrote for the human. And the ones who figured out how to do both are running marketing organizations right now.

This is that moment, again. The robot has gotten dramatically more sophisticated — it’s not crawling for keywords anymore, it’s inferring purchase decisions — but the fundamental dynamic is the same. A new machine is in the distribution chain. It has preferences. Those preferences are legible. You can learn them.

Attribute completeness — making sure your product data is complete, structured, validated, and machine-readable at every layer — is not a technical task you hand off to your dev team and forget. It’s a brand discipline. It’s the new SEO, with the same mixture of craft and rigor that SEO required when it actually mattered.

The marketers who treat this as a crisis are the ones who only knew how to do one thing — tell stories to humans. That’s not nothing. It’s foundational. It still matters enormously. But one tool is a fragile practice. The marketers who treat this as an expansion are building the next version of the discipline — one that requires understanding how meaning travels through systems that include machines, not just people.

Your brand just became a spreadsheet. That’s true. And the best-performing spreadsheets are the ones with the most complete data.

That’s not a demotion. That’s just the job now.

This is not a technology story. It’s a career story.

Your email team celebrated a 48% open rate last quarter. Someone put a slide together. There were probably high-fives. But a share of those opens — a share that’s growing fast and that nobody on the team is measuring — were AI agents. Background processes scanning inboxes on behalf of humans who never saw the message, never saw the subject line, never registered your brand at a conscious level. Your team got high-fived for impressing a robot.

Here’s what makes that a career problem, not just a data problem: the thing your job is measured by is increasingly a fiction. You optimize for the score. The score stops meaning what it used to mean. But the raise still depends on the score.

And nobody is decomposing the signal. Not your ESP. Not your CRM. Not your analytics stack. The number goes into the weekly business review as if it means one thing, when it actually means two very different things layered on top of each other — and only one of them matters for persuasion.

The Ghost Click Problem

Here’s what’s actually happening inside your dashboard.

A growing class of “recipient agents” — tools like OpenClaw (formerly Clawdbot/Moltbot) and similar inbox automation layers — monitor email inboxes proactively. They crawl and summarize incoming messages every 15 to 30 minutes. When one of these agents “reads” your email, it triggers your tracking pixel. It downloads your assets. It registers in Klaviyo or Salesforce as a human open. The signal looks identical. You can’t tell the difference. Most teams aren’t even looking for it.

This isn’t a fringe problem. The 2025 Imperva Bad Bot Report found that automated traffic surpassed human activity for the first time in a decade — bots now account for 51% of all internet traffic. Akamai reported AI bot traffic across their network surged more than 300% over 2025. The inbox is just one place this is showing up.

Then there are the ghost clicks. Autonomous agents use automated browsers to follow links, verify product availability, compare competitive pricing, scrape structured data. The user never sees the landing page. The agent does. Your CTR goes up. Your conversion rate doesn’t move. You chalk it up to a weak landing page and start another A/B test.

Meanwhile, something else is happening to the creative work. The layout you labored over, the emotive headline, the brand imagery your designer spent two weeks on — the agent’s summarization engine strips all of it. The human user gets a text-based distillation of the core offer. Subject line, discount, expiration date. That’s what survives the translation. Everything else was for an audience that was never reading.

I wrote about agentic commerce last month — the shift from human-initiated transactions to agent-completed ones. This is what that shift looks like from inside your dashboard. It looks like great numbers.

What Your Org Chart Assumes

Most marketing teams are built on a premise that made sense five years ago: different humans use different channels, so different specialists optimize each channel for human engagement.

The email team owns opens, clicks, and revenue attribution. The content team owns traffic, time-on-page, and SEO ranking. The paid team owns CTR, ROAS, and conversion rate. Each team has a scorecard. Each scorecard drives headcount decisions, performance reviews, and promotion criteria.

The org chart assumes the signals are clean. That an open is a human opening. That a click is a human clicking. That time-on-page is a human reading.

Those assumptions are eroding — not all at once, not uniformly, but steadily and without announcement. I’ve watched it happen. The erosion is faster in B2B, where procurement teams have been using research agents and competitive intelligence tools for years. It’s accelerating in B2C as inbox automation becomes a consumer product rather than an enterprise one. And nobody sends a memo when the signal breaks.

The email manager isn’t managing a human relationship anymore. She’s managing machine readability, list hygiene against bot traffic, and deliverability signals in an environment where inbox agents are the first filter — not the spam folder. She doesn’t know this, because nothing in her dashboard has been updated to reflect it. Her KPIs were written when the audience was entirely human. They haven’t been revised.

That’s the org chart problem. It isn’t that the roles are wrong. It’s that the roles are oriented around a measurement system that is quietly describing a different reality than the one it’s reporting on.

It’s a New Scoreboard

There is a new set of metrics emerging for a world where agents are a primary audience. The terminology is still settling — the field is moving that fast. But two concepts matter most right now.

Inference Density* How efficiently does a brand’s content get ingested by an agent’s reasoning engine? Not word count — utility per token. A well-structured product page with clear specifications, schema markup, and clean data architecture scores high. A page built for emotional resonance and visual impact scores low. This is the new “readability score” — except the reader is a language model deciding whether your brand is worth citing.

Citation Prominence* As AI answer engines become a primary research surface — Gartner projects traditional search volume will drop 25% by 2026 — being the cited source, not just the ranked result, becomes the actual prize. Answer Engine Optimization isn’t a niche anymore. It’s the emerging discipline underneath SEO. The marketer who can engineer her brand into the citation layer of AI-generated answers is doing something fundamentally different from the marketer optimizing for position-three Google rankings.

Neither of these is on any marketing team’s current dashboard. They should be.

And here is where I want to be direct about the career stakes, because that’s actually the story.

I have spent years in rooms where the scoreboard was open rates, click-through rates, revenue attribution by channel. I know what it feels like to build a career on those numbers. I also know — because I’m watching it happen — that the ground underneath those numbers is shifting. Not slowly. Not theoretically. Now.

The marketer who can speak fluently to inference density and citation prominence in a QBR is not describing the same job as the marketer running subject line A/B tests for open rate. They are different jobs. One of them has a future that compounds. The other one — and I say this as someone who has run those tests, who has celebrated those open rates — is optimizing for a scoreboard that is becoming decorative.

This is not about learning to code or becoming “technical.” It’s about understanding what’s actually happening to the audience your work is supposed to reach. The marketers who recognize this shift early are the ones who will write the next round of job descriptions rather than respond to them. That’s a different kind of leverage. It determines whether you’re the person defining what the team measures, or the person being measured by metrics someone else chose — metrics that may not mean what they used to mean.

That leverage is available right now. To anyone paying attention.

Who Gets There First

The marketers who figure this out first won’t announce it on LinkedIn with a hot take. They’ll quietly reframe their team’s measurement system. They’ll start asking their ESP for raw data on open timing and location signals — looking for the bot-pattern fingerprints. They’ll push for structured data audits. They’ll start running experiments on information density rather than emotional resonance and watching what actually correlates with human conversion downstream.

They’ll get weird looks in the weeklies. And then, about eighteen months from now, someone will write a case study about what they did, and everyone will nod and say they saw it coming.

The metrics we’ve been given are not neutral instruments. They are artifacts of a specific era — the era of humans as the primary audience for marketing signals. That era is not over. But it’s no longer the whole story, and it hasn’t been for a while.

Your 48% open rate is a composite signal. Part human, part machine, indistinguishable in your current stack. Your team celebrated a number. The question worth asking — the one that your next performance review will not ask, but probably should — is: who were you actually talking to?

The marketers who can answer that question, and build toward a cleaner answer, are the ones doing the most important work in the field right now.

They’re also — not coincidentally — the ones who will still have jobs worth having when the dust settles.

Someone DM’d me a “skills pack” last week. An AI marketing toolkit — free download. Every file was under a hundred lines. No executable code. Articles about what Claude could do *if* the skill existed, packaged as if they’d work right out of the box. They don’t.

Is it a scam? Maybe. It’s definitely a symptom. The same distance between what’s being pitched and what’s actually inside the box is playing out across the entire marketing technology landscape — from free downloads in your DMs to a market correction that just wiped out the equivalent of the GDP of Finland.

The Model Does the Work. The Wrapper Takes the Credit.

I sit inside these platforms every day. I’m working with in-house marketing teams and agencies — building, auditing, consulting — and the pattern is impossible to miss once you’re looking at the architecture layer. You can’t see it from a spreadsheet. You see it when you open the box.

A CMO I was speaking with this week asked me to look at an LTV prediction tool her team had been pitched. The vendor was quoting them roughly $75,000 a year. She pulled up the demo and asked: what do you think? Is something like this worth it?

It wasn’t. Not at a glance, and not after digging in.

The platform they’re already on — Klaviyo — does all of this natively. LTV prediction, churn risk, send time optimization, audience building from natural language prompts, and as of last fall, autonomous Marketing Agents that plan, create, and optimize campaigns (with human guardrails, of course). The $75K tool was selling back functionality their ESP already gives them, just wrapped in a different UI, positioned as something proprietary, priced like it was irreplaceable.

This isn’t a one-off. The martech landscape hit 14,106 solutions last year — over 3,000 new products in twelve months, most of them wrappers repackaging a foundation model with a dashboard and a subscription fee. A meaningful number of the tools that emerged around Klaviyo, Shopify, Meta, and Google are reselling capabilities the platforms already ship natively. They just have better landing pages.

Paid media is worse. A whole category of startups and AI-native agencies have promised to “beat the algorithm” — their AI would outsmart Meta’s AI. Except Meta’s Advantage+ campaigns use deep learning across 3+ billion users to optimize targeting and creative in real time. Third-party tools can’t access that signal data. They can’t see the auction. They’re sitting outside the house, guessing what’s happening inside, and charging you a monthly fee for the guesses. Advantage+ generates 22% higher ROAS than manually optimized campaigns. The tools claiming to beat the platform are underperforming its default setting.

To be clear: some third-party tools earn their keep. The ones that bring proprietary data, that layer on creative intelligence the platform doesn’t surface, that solve a genuine workflow gap — those are real products. But the bar for “genuine” is higher than most vendors want you to think, and the buyer usually can’t tell the difference. That’s the problem.

Nobody Had Time to Vet. So the Stack Grew.

The wrapper problem isn’t just a supply-side issue. It’s a literacy problem — and an operational one.

I consult with a brand whose martech stack had ballooned to the point where they were paying for something like 54 tools and plugins, using fractions of each, and none of them were fully implemented. When I audited just their ESP, I found 11 active integrations feeding into it — and five of them weren’t even being used. The flow sprawl was worse: 124 automations in the account, only 30 actually live, with seven-plus versions of the same abandoned cart flow sitting in draft because nobody had archived the old ones. The monthly spend across the stack ran well into five figures, and the overlap was invisible to everyone except the person I was brought in to be.

The team wasn’t lazy. They were drowning. BAU demands — campaigns to ship, reports to pull, fires to put out, and the meetings (the meetings!) — it consumed every hour. The executive suite was simultaneously pressuring them to “show results this month” and “adopt AI.” There was no time, no bandwidth, and honestly no structural support for anyone on that team to sit down and evaluate whether the tools they were paying for duplicated what their core platforms already did. Nobody had the space to vet. So the stack grew, relatively unchecked, for *years*. The overlap compounded and the budget bled.

I want to be clear: This isn’t a failure of the people. It’s a failure of the system they and so many teams are operating in.

The data confirms it. A Hootsuite study found that 64% of marketing leaders believed their AI tools used real-time data. Only 39% of practitioners — the people actually inside the tools — agreed. That 25-point gap is the distance between what leadership thinks they bought and what the team knows they’re using.

The technical literacy required to evaluate these tools — to distinguish between genuine machine learning and an if/then script with “AI” in the name — was never part of the marketing career path. Most marketing leaders came up through brand, performance, or strategy tracks. Nobody taught them to read an API call. And the vendors are counting on it.

Gartner found that 64% of CMOs believe they lack the budget to execute their strategy — and yet 65% believe AI will dramatically change their role. Leaders who feel under-resourced and existentially threatened will buy anything that promises to close the gap. The vendors aren’t selling software. They’re selling relief from the anxiety of being left behind.

The industry has a word for this: shelfware.

I’d call it a fear tax.

The Invoice Has Arrived.

The bill has come due. Just this month, roughly $300 billion in market value evaporated from the software sector — investors realized that generative AI models were becoming substitutes for SaaS features, not just enhancements. Analysts called it the “SaaSpocalypse.” An MIT report found that 95% of corporate AI pilots never reached production. Forrester predicted enterprises would delay a quarter of planned AI spend into 2027 because nobody could demonstrate the ROI.

Some of it was outright fraud, sure. The FTC launched “Operation AI Comply,” suing companies like Ascend Ecom and Growth Cave for marketing “AI-powered” business tools that were either nonexistent or required the same manual work the buyer was trying to escape. When “AI” is a magic word you can attach to anything and nobody in the room can verify the claim, this is where it ends.

But for the teams like the one I described — the ones already stretched thin, already paying for tools they’re barely using — this correction isn’t abstract. It’s the moment someone finally looks at the credit card statement and asks what they’re actually getting for the money. And I suspect, nobody on the team really knows.

What’s Actually in the Box.

I build marketing infrastructure with AI. Not in the abstract — I write the agents, deploy the servers, build the automation pipelines, and work inside the platforms daily. Klaviyo, Meta, Shopify, analytics stacks. The orchestration layer underneath.

When you work at that level, you develop a kind of X-ray vision for what’s real and what’s a forwarding address. Not because you’re smarter — because you’ve opened enough boxes and spent meaningful time in them.

The CMO who asked me about the $75K LTV tool? She’s sharp. She runs a strong organization. She just hadn’t had a reason to look at what was underneath a vendor’s demo before. Why would she? She’s never had to. That wasn’t her job, and may still not be her job — but the market has made it everyone’s job.

Here’s what I’ve learned to ask.

Does the tool have data the platform doesn’t? If it’s “optimizing” your Meta campaigns but can’t access Meta’s auction data without latency, it’s guessing. If its only data source is the same API you can already access, the value proposition is the pretty dashboard — not the intelligence.

Does it execute, or does it just recommend? There’s a meaningful difference between a tool that writes back to the ad platform and one that generates a PDF of suggestions. True automation closes the loop. A recommendation engine with a subscription fee is an agency you can’t fire.

Is the platform shipping this feature next quarter? Klaviyo’s Marketing Agents made dozens of email drafting tools obsolete overnight. If you’re buying a capability the platform is actively building into its core product, you’re renting a feature with an expiration date. Gatsby was a genuinely good product, so Klaviyo bought it. Now it’s native.

Many marketing teams can ask these questions, but they can’t vet the answers. Not because they’re incapable, but because they’re buried in execution and nobody built this into the evaluation process. That’s how we ended up here.

The Point.

The skills pack in my DMs wasn’t a scam. I’m choosing to believe that. It was someone who may genuinely believe she’s offering a useful product — and couldn’t tell the difference between describing a capability and building one.

The $75K LTV tool probably isn’t a scam either. It’s a real product, with a real sales team, quoting a real price for something the client’s existing platform already does.

The tool bloat for my client wasn’t a failure of intelligence or effort. It was a brilliant team doing their best inside a system that doesn’t give them the time or the framework to evaluate what they’re buying.

That gap — between describing and building, between the pitch and the product, between what leadership thinks they bought and what the team knows they’re using — is the wrapper problem. It exists at every scale, from a free download in your DMs to a $300 billion market correction.

The tools that survive will be the ones that have proprietary data, that execute rather than recommend, and that do something the platform can’t do natively. Everything else is a just wrapper with a landing page.

The correction isn’t a catastrophe, clearly. It’s just a filter. And the filter is going to reward the same thing it always has — the people who actually build, who can show the architecture, who know what’s inside the box because they put it there.

If you’re running a marketing team right now and you’re not sure what is or isn’t snake oil, here’s where I’d start:

Audit the overlap* Pull every tool in your stack and map it against what your core platforms already do natively. You will find redundancy. The brand I audited had five unused integrations and 94 dead automations — and nobody knew until someone looked.

Ask the three questions before you buy anything new* Does it have data the platform doesn’t? Does it execute or just recommend? Is the platform shipping this feature next quarter? If you can’t get clear answers, walk.

Give someone the time to vet* This is the one nobody wants to hear. Evaluation takes bandwidth your team doesn’t have — which is exactly why the stack grew unchecked in the first place. Build it into the process or accept that you’re paying the tax.

Read the platform roadmap* Klaviyo, Shopify, Meta, Google — they all publish what’s coming. If the tool you’re evaluating solves a problem the platform is actively building into its core product, you’re renting a feature with an expiration date.

None of this requires a consultant. It requires someone on your team with the time and space to learn.

Subscribe now

I’m bad at LinkedIn. I know this about myself. But one of my goals for 2026 was to find my voice here, and I’m currently climbing cringe mountain to do it.

I’m a data-driven person, always have been. I measure everything. So naturally, I built something to measure this too. It turns out LinkedIn knows a lot more about me than I realized — and much more than it tells me.

“It’s Really Difficult to Do This Yourself”

Last week, I was reeling a bit after the third call with a career coach disguised as a recruiter in as many days. A line from one conversation kept echoing through my head: “it’s really difficult to do this yourself.”

Ok, girl. Bet.

So I went deep. The kind of deep where you stare at your own LinkedIn profile until the words stop meaning anything. Five rounds of headline optimization. Version four of my About section. You know the mode. The one where you’re rewriting your career story in 2,000 characters and wondering if “P&L Owner” or “Revenue Leader” is the keyword that gets you past the algorithm.

Somewhere in that process, I requested my LinkedIn data export. Not for any strategic reason. Just curiosity — what does LinkedIn actually have on me?

What came back was a zip file. Seven CSVs. Connections. Messages. Invitations. Company follows. Shares. And two files I didn’t expect: one called `Inferences_about_you.csv` and another called `Ad_Targeting.csv`.

LinkedIn has been making inferences about me — guesses about my demographics, interests, and behavior — and, of course, selling access to those guesses to advertisers. I knew this intellectually. Seeing the actual file was different. There were hundreds of entries. Some accurate, some wildly wrong, all of them opaque. LinkedIn never showed me this. I had to go find it.

Raw Data, Few Answers

The CSVs were dense and, on their own, useless. Raw data without structure. So I did what I’ve been doing a lot lately — I built something.

I wrote a tool that takes your LinkedIn data export and generates 10 interactive visualizations — network maps, message analysis, career timelines, even a comparison of what LinkedIn *thinks* it knows about you versus what your data actually shows. All rendered in standalone HTML files you can open in any browser. No server. No installs. No data leaves your machine.

It took just an hour or so. And it told me more about my professional life than 14 years of using LinkedIn ever did.

How Big Is Your Network, Really?

The visualization that hit hardest was Connection Quality. It cross-references every one of your connections against your message history and sorts them into three buckets: Active (you’ve talked recently), Dormant (you talked once, years ago), and Never Messaged.

The “Never Messaged” number is the honest answer to “how big is my network, really?”

Mine was humbling. I won’t share the exact breakdown, but I will say this: the number LinkedIn puts at the top of your profile and the number of people you’ve actually had a conversation with are two very different things. We all know this intuitively. Seeing it quantified is something else.

A Career in Strata

The other one that was fun to sit with was Career Strata. It takes your entire connection history, divides it into phases based on when you connected with people, and stacks them like geological layers. Oldest at the bottom. Newest at the top.

My deepest layer is labeled “Bedrock (2012–2014).” College. The Apple Store in Grand Central. The Tumblr community that started everything.

Above that: The Line, where I went from intern to director. Then Snowe, Paravel, Brooks Brothers. The consulting years. Net32. Each layer tells a story — not through job titles or bullet points, but through the people I was connecting with during each chapter. Where the network was growing. Where I was heads-down and not reaching out at all.

It’s your career, told through the people in it. LinkedIn has had this data for over a decade. They never built this view.

You Are the Product and the Last to Know

One visualization compares LinkedIn’s algorithmic inferences against your actual data. For each guess LinkedIn has made about you, it pulls what your export actually shows and renders a verdict: Accurate, Partial, or Wrong.

Some of it was fine. Some of it was laughably off. LinkedIn has assigned me 739 skills — including Laser Physics and Road Transport. It filed me under 161 interest categories, among them Robotics and Insurance. It targets me to advertisers as a Technical Support Specialist and a Sales Lead. I’m a fractional CMO who’s spent 12 years in growth marketing. But sure, laser physics.

The point is that LinkedIn is making hundreds of assumptions about who I am, selling access to those assumptions, and never once asking me if they’re right.

This is the data asymmetry that I found most interesting. It’s not that platforms collect data. That ship sailed. It’s that they collect it, analyze it, monetize it, and then hand you back a connection count and a notification that someone viewed your profile. The intelligence flows one direction.

The Questions LinkedIn Won’t Answer

Some of the visualizations answered questions I didn’t know I had. One charts posting frequency against new connection requests. The correlation of 0.91 — nice. In months where I posted consistently, I got 1,765% more new connections than in months where I didn’t. Correlation isn’t causation. But it’s a hell of a data point for someone who used to post once a year.

There are more questions in that data than I’ve built visualizations for. What types of messages actually get responses? Which connections lead to real professional relationships versus permanent dormancy? Is there a correlation between when you connect with someone and whether you ever speak again?

LinkedIn has the data to answer all of these. They just don’t.

The Audacity to Bootstrap

Last fall I wrote about vibe coding and the personal software revolution — the idea that the most interesting thing about AI-assisted development isn’t building the next App Store hit, but building personal tools for an audience of one.

This tool is that idea, made a little more concrete.

I didn’t build this to sell anything. I didn’t build it to launch a startup. I built it because I had a question — what can Linkedin actually do for me? — and the platform didn’t want to answer it. So I built the answer myself, and I built it for you, if you’d like to try it.

Check out the demo

To the career coaches sliding in my inbox, attempting to sow doubt, attempting to trigger my loss aversion, and create new insecurities so I’ll hand over a few thousand dollars — I see you.

Yes, it may be difficult to do this myself. Maybe I’ll fail and end up crawling back to one of you.

But what if it works out?

I’ll just keep building and let you know.

Subscribe now

Everyone wants a “storyteller” now.

Not a content marketer. Not a copywriter. A *storyteller*.

LinkedIn job postings using the term have doubled in the past year. Over 50,000 marketing positions now include it. Executive mentions of “storytelling” on earnings calls jumped from 147 in 2015 to 469 in 2025. Microsoft, Google, OpenAI — they’re all hiring “chief storytellers” and paying upwards of $274,000 for the privilege.

Here’s my question: storytellers to tell *what*, exactly?

Most Companies Don’t Have a Story Worth Telling

They have a product. They have features. They have a mission statement someone workshopped in an executive offsite three years ago. But a story? Something that actually means something to anyone outside the building?

Not so much.

The storyteller gold rush isn’t about companies suddenly discovering the power of narrative (right? right). It’s about companies hoping they can hire their way out of being boring.

It doesn’t really work like that.

What “Hire a Storyteller” Usually Means

When a company posts a storyteller role, what they’re often really saying is:

”We don’t have a clear point of view, and we’re hoping someone can create one for us.” A storyteller can’t invent your company’s reason for existing. If leadership can’t articulate why you matter, no amount of narrative polish will fix that.

”Our content isn’t performing, and we think better writing is the answer.” Sometimes it is. More often, the problem is that you’re saying the same thing as everyone else in your category, just with different fonts.

”We saw that other companies are doing this and we don’t want to be left behind.” The best reason to hire anyone, obviously.

*To be clear: storytelling skills are genuinely valuable. Story-driven content improves recall by 22 times. Brands with compelling narratives see 20-30% higher customer loyalty. In a world where 70% of web content is now AI-generated slop, authentic human voice matters more than ever. A great storyteller can absolutely help a brand find its voice, shape its narrative, and connect with customers in ways that resonate.

But storytelling is a multiplier, not a magic wand. The company has to be willing to meet the halfway — to have a perspective, to hold an opinion, to say something that might not land with everyone. You can’t show up in public discourse in a way that reverberates, that speaks to real emotion (and not just rage), if there’s no substance underneath. A storyteller can shape and amplify but very rarely can they manufacture meaning from nothing.

What Actually Makes a Story Worth Telling

The companies with stories worth telling share a few things:

There must be a genuine point of view. Not a positioning statement. An actual opinion about how things should be. Something that might piss some people off — I’d just recommend pissing them off for the right reasons. If everyone agrees with your “brand story,” it’s not a story. It’s wallpaper.

The stakes have to actually matter. Stories require tension. What are you fighting against? What’s at risk? “We want to help businesses succeed” isn’t a story. “We believe the way enterprises buy software is fundamentally broken and we’re going to fix it” — that’s the beginning of one.

The actions have to match the words. The fastest way to kill a brand story is to tell one thing and do another (ahem). Customers aren’t stupid. They know when you’re lying and can smell the gap between your narrative and your reality from a mile away.

The story needs to exist independent of marketing. The best brand stories aren’t invented by marketing departments or your branding agency. They’re part of the PMF, they’re discovered, they’re already happening in how the company operates, treats customers, makes decisions. Marketing’s job is to find that story and tell it well — not to fabricate one from scratch (see point #1 and start again).

The Uncomfortable Truth

The storyteller gold rush may create a lot of jobs. Many of them will be frustrating as hell.

Because talented writers and narrative strategists will join companies expecting to tell meaningful stories, only to discover there’s no story there to tell. They’ll be asked to spin straw into gold, to make undifferentiated products sound differentiated, to manufacture authenticity where none exists.

Some will succeed in spite of this — they’ll dig until they find something real, or they’ll push the company to actually develop a point of view worth sharing.

Most will burn out, churn out, or learn to produce the same forgettable content as everyone else, just with better prose.

If you’re a storyteller considering one of these roles, do your due diligence. Make sure there is product-market fit. Ask hard questions in the interview. Does this company actually have something to say? Do the leaders have a genuine point of view? Is there substance here, or just a desire for the appearance of substance?

And if you’re a company hiring storytellers, maybe start by asking whether you have a story worth telling in the first place. If yes, are you willing to invest in them?

I restarted my consulting and advisory practice ten months ago. I’m having a lot of fun, learning a lot, and the time-boxing of consulting work allows for time to dedicate to continuing education, networking, certifications, writing, and developing new skills — something I believe is critical in this new age of AI that is rapidly changing the way we approach work. I’m lucky enough to have my practice while I have conversations and look for my next full-time role. The option to be picky is a luxury, and it’s not one I take for granted.

About six months ago, I joined Chief (yes, I know, that’s a topic for another time) but since networking has never come naturally to me, I determined it was a sound investment. Upleveling and all that. Fast forward to a few weeks ago, I put out a call to my other Chief members for recommendations on a career coach — I’ve always had difficulty reaching out for help, and I thought, naively or not, that Chief would be a decent place to start. A few names were shared with me, and it’s an interaction with one that inspired this post.

I connected with this coach on LinkedIn, sent the “Hi, I was referred to you” email, explained a little about what I was hoping to accomplish, and attached my resume.

Normal, right?

The response I got was... interesting.

Something felt off, so I declined, and it was the response that I got back confirmed every instinct I had.

The blatant high-pressure sales tactic, the qualifying out, the artificial scarcity and exclusivity, the attempt at triggering loss aversion, and topped off with some manipulative “priorities” framing. The implication that if I don’t buy, I don’t value my career or growth or future. A conspicuous guilt lever disguised as insight.

Girl, ew.

Noticeably absent was any attempt to understand my situation. No demonstration of her relevant expertise or explanation of what her ‘diagnostic’ fee actually buys. The attempt to make me feel like I’m not good enough for the cool kids’ table... yet. A particular brand of negging I have never experienced in a professional setting and never would have expected from a referral in a women’s leadership group, member to member.

Here’s the thing: I’m not writing this because I fell for it. Her methods certainly worked, because I’m feeling a certain way, but I’m writing this because I know how many people might, especially in a job market as opaque and volatile as the one we’re currently in.

The Red Flags

Let’s start with the facts she conveniently ignored — facts I’d like to remind you of, dear reader:

The average executive searches are *long*, and that’s completely normal. Multiple executive search firms report that the average executive search lasts roughly 10 months. VP roles take 4-6 months minimum. C-suite can stretch to a year or more.

So when someone tells you that “most senior leaders aren’t blocked by effort” after three months, 6 months, or 10 — as if you’re doing something wrong — they’re either uninformed or manipulating you. Pick which you think is more likely.

There are also a few high-pressure sales tricks you’ll want to keep an eye out for, either from gurus wanting to sell you a program or in your next RFP process with a marketing agency:

1. Diagnosing problems without information. “Unclear decision structure” and “misread market signals” based on a three-sentence email? That’s not insight. That’s a cold read designed to make you doubt yourself, don’t fall for it.

2. Dismissing support as worthless. “Those tend to feel supportive without changing outcomes.” Translation: emotional support and human connection don’t matter. Only my $2,500 diagnostic matters.

3. All-or-nothing gatekeeping. “I don’t do informal meetups.” Legitimate coaches offer consultations. They want to help you, not just qualify you as a lead.

4. Scarcity as a feature. “My work isn’t designed to be accessible to everyone.” This is positioning exclusivity as value. It’s manipulation dressed up as standards.

She’s Not the First

This coach isn’t an outlier. She’s the third or fourth I’ve encountered in the past year that operates this way — and those are just the ones who came through trusted referrals. My LinkedIn inbox is filled with outreach from others posing as recruiters that are really just gurus looking to sell me their coaching package.

I get it, everyone’s gotta eat, but there’s no reason to be so icky about it.

The executive career coaching industry has a predatory underbelly that specifically targets senior job seekers. The playbook is consistent:

1. The fees are staggering. We’re not talking $200 for a resume review. Executive coaching engagements routinely run $5,000 to $10,000 upfront, with some packages hitting $13,000 to $30,000 for six months. My $2,500 “diagnostic” was the entry-level shakedown.

2. The outcomes aren’t guaranteed — and can’t be. Only employers make job offers. No coach controls that. Yet they sell “60-90 day timelines” and promises of “executive presence” and “personal brand” as if these are deliverables. They’re not. They’re vibes dressed up as methodology.

3. The systems are deliberately vague. Ask what you’re actually getting for $10,000 and you’ll hear about “structured diagnostics,” “repositioning frameworks,” and “market signal alignment.” Push harder and you get nothing concrete — because there’s nothing concrete to give. The ambiguity is the product.

4. They target vulnerability by design. If you’ve been searching for months, you’re already questioning yourself. Predatory coaches exploit that — using your fear and uncertainty to justify fees that would make you pause if you were employed and confident. One documented case describes a coach quoting $4,000 upfront and demanding immediate commitment, insisting that people who “don’t commit right away usually don’t follow through.” That’s not coaching. That’s a high-pressure sales close.

Remember, the career coaching industry is unregulated.No standardized licensing. No oversight. No accountability. Anyone can hang a shingle and charge whatever the market will bear — and when the market is desperate senior executives who haven’t job searched in years, the market will bear a lot.

If you’re months into a search, you don’t need someone to manufacture new insecurities. You need tactical feedback, peer support, and maybe a therapist who won’t also try to upsell you.

What Actually Helps

Peer support. Other executives in the trenches — whether through communities like Chief or Pavilion or ExecThread or Exit5, a Slack group of former colleagues, or just a few people you trust enough to be honest with. “Yeah, this sucks, and it’s taking me a long time too” is more valuable than any one framework.

Specific feedback. Not vague diagnoses — actual input. A recruiter who’ll tell you what’s not landing. A former colleague who’ll run a mock interview. A trusted peer who’ll say “that story doesn’t make you sound like the leader you are.” Real data, not vibes.

Recruiters who work your level. Executive search firms and retained recruiters understand market signals, compensation benchmarks, and hiring timelines better than any coach who’s never placed someone in your role. Cultivate those relationships.

Free and affordable resources. AI-powered interview prep tools like Superinterviews, networking tools like Happenstance and Clay.Earth.

The Bottom Line

I’m not saying all career coaches are the bad guys out to scam you, but if someone makes you feel small to sell you something, they’re not worth your time.

You’re not blocked by “unclear decision structure.” You’re navigating a tough market with limited roles, long timelines, and a thousand variables outside your control.

Don’t let anyone use your vulnerability against you.

Have you encountered predatory coaching tactics in your job search? I’d love to hear your stories.

Leave a comment

Subscribe now

Three years ago, I was sitting on a roundtable at ShopTalk when the conversation turned to AI. ChatGPT had launched about four months earlier, and the energy in the room was a particular mix of excitement and uncertainty that I suspect most of us remember. Everyone was trying to figure out what generative AI meant for retail, for marketing, and for the way we work.

I remember saying something that felt, at the time, almost too obvious to be worth saying out loud: the output is only as good as the input, and the prompt is only as good as the data your AI is being trained on.

As I prepare for ShopTalk 2026, I keep returning to that conversation. Not because I was prescient — I’m not, clearly — but because three years of exponential progress in AI capabilities have done remarkably little to change the underlying constraint.

The Acceleration We Expected (and the One We Didn’t)

The technology has evolved faster than most of us expected. We’ve moved from general-purpose chatbots to specialized agents with domain-specific capabilities within a matter of months. From tools that answer questions to systems that execute multi-step workflows autonomously, and quite gracefully. The gap between “impressive demo” and “production-ready application” has narrowed considerably, and it’s only going to shrink further.

What I didn’t fully appreciate in 2023 was the velocity of that acceleration. While we were still laughing at Will Smith eating spaghetti, each capability was unlocking the next. Agents that can reason about code enable agents that can build, QA, and deploy. Agents that can navigate interfaces enable agents that can operate across systems.

What’s become clearer with time is the ceiling on what AI can do for any given organization isn’t primarily a function of model capability. It’s a function of data architecture.

I said it then and I will say it again now:

If your data is shit, if your input is trash, don’t be shocked if your AI can only parse paper from plastic.

Excuse my French.

The Taxonomy Problem

I’ve worked with companies that have over 100,000 SKUs, thousands of brands, hundreds of manufacturers, and a category taxonomy that was largely the brainchild of a long-absent former employee. These systems work, in the sense that products get listed and orders get fulfilled. But they work through institutional knowledge, human pattern-matching based on personal belief systems, and a kind of organizational muscle memory that doesn’t transfer well to automated or autonomous systems.

Ask a new employee to navigate that taxonomy on their own and they’re lost for about 6 months. Ask an AI agent to optimize that catalog — to surface the right products to the right customers, to identify gaps in the assortment, to automate merchandising decisions, what have you — and you’re asking it to reason with a structure that was never designed to be reasoned with by anyone other than the people who built it.

Don’t take it personally; taxonomies evolve organically. Categories that made sense when you had 10,000 SKUs don’t scale cleanly to 50,000 or 100,000. The sub-sub-sub-category that captures an important distinction for one product manager is meaningless noise to another. Entropy is the natural state of any sufficiently large data system.

But the cost of that entropy is changing. What used to be friction — extra clicks, occasional mis-categorizations, the need for tribal knowledge — is becoming a hard ceiling on what’s possible.

Specialized Agents and Their Dependencies

The trajectory of AI development points clearly toward specialization. You know this. I know this. General-purpose models will continue to improve, but the real leverage comes from agents designed for specific domains: merchandising agents, pricing agents, creative development agents, analytics agents. Systems that don’t just know things in general but know how to operate within a particular business context.

A specialized agent that understands the nuances of e-commerce product data is more valuable than a general-purpose model that knows a little about everything. Though the distinction between “specialized model” and “general model acting specialized” is blurring slightly, if we’ve learned anything, it’s that domain expertise matters as much in AI systems as it does in human teams.

But specialization creates a different kind of dependency. A merchandising agent needs clean product data and context and category knowledge to merchandise effectively. A personalization agent needs consistent customer data to personalize meaningfully. An analytics agent needs well-structured event data to generate insights worth acting on. The more capable the agent, the more it exposes the gaps in the data it’s attempting to reason through.

The Boring Work

There’s nothing glamorous about data hygiene. Taxonomy audits don’t make for exciting Linkedin posts. Documentation reviews don’t generate conference buzz. The work of maintaining clean, consistent, well-structured data is precisely the kind of work that gets deprioritized when there are shinier problems to solve, and I’ve witnessed it firsthand. Throughout my career as a marketer, I have been in heated conversations with peers in Engineering who didn’t see the value in custom event parameters or granular profile dimensions or why that should jockey for position on the roadmap against some other sexier feature.

I’ve noticed that the organizations making real progress with AI tend to have something in common: they did the boring work first. Not perfectly — because no one’s data is perfect — but well enough that automated systems can reason through it without constant human intervention.

Clean taxonomies. Consistent product attributes. Regular hygiene processes. Documentation that doesn’t live exclusively in one person’s head. These aren’t AI initiatives, but they’re the foundation that makes AI initiatives possible.

The Question I Keep Asking

As the technology accelerates, the gap between organizations that can leverage it and those that can’t is widening. The limiting factor isn’t access to models; those are increasingly commoditized. It’s not even technical and strategic talent (though that matters). It’s the accumulated decisions about data architecture made over years or decades, and the willingness to revisit those decisions in light of what’s now possible.

So the question I asked in 2023 is the same question I’m asking now, with more urgency: Is your data ready for what’s coming next?

I imagine for most organizations, the answer is probably not. But the organizations that start addressing that gap now will have a meaningful advantage over those that wait until the bottleneck becomes impossible to ignore.

Leave a comment

Let’s get something out of the way first: if your “employee advocacy program” involves a Slack channel where marketing drops pre-written LinkedIn posts and asks everyone to hit share, you don’t have a strategy. You have a cargo cult. And everyone — your employees, your audience, the algorithm — can smell it.

I’ve sat in the meetings where these programs get pitched. The deck always has the same slide: “Employees have 10x the reach of brand channels!” And technically, that’s not wrong. LinkedIn’s own research shows content shared by employees receives eight times more engagement than content pushed through the company page. The Hinge Research Institute found that leads from employee advocacy convert at seven times the rate of other leads.

But here’s what the deck never addresses: why would anyone actually do this?

Not the company. The person. The engineer who’s already underwater on a sprint. The customer success manager juggling fifteen accounts. The designer who’s three revisions deep on a rebrand that keeps shape-shifting. Why would they add “become a LinkedIn thought leader” to their list?

The answer can’t be “because it helps the brand.” That’s not a reason. That’s a mandate dressed up as an opportunity.

The programs that actually work — and I’ve seen a few — start from a completely different premise. They start by asking: Who on this team already wants to build something for themselves?

Not everyone does. That’s fine. Most people want to do their job well, go home, and not think about personal branding. Respect that. But in any organization of reasonable size, you’ll find a handful of people who light up at the idea of sharing what they know. They’re already posting occasionally. They’ve mentioned wanting to write more, or they’ve asked about speaking at a conference, or they have opinions about their craft that they’re dying to articulate.

Those are your people. And your job isn’t to hand them a content calendar. It’s to get out of their way while giving them the support they actually need.

That support looks like access — to data they can reference, to leadership for quotes, to design resources so their posts don’t look like they were made in PowerPoint on a plane. It looks like a legal and PR review process that moves in hours, not weeks, because nothing kills momentum like waiting nine days for approval on a LinkedIn post about API architecture. And it looks like genuine investment in their growth: training on how to write well, how to structure a video, how to build an audience that matters to them, not just to the company’s pipeline.

The transaction has to be honest. You’re asking someone to put their name and face on content that benefits your brand. In return, they should get something real: visibility in their industry, a network that follows them if they leave, skills that compound over their career. If the only winner is the company, you’ll get compliance at best and resentment at worst.

I think about Lovable, the AI development platform, and what they’ve built in public. Anton Osika, the company’s co-founder and CEO, shares technical deep-dives and honest reflections on what’s hard about AI development — the unsexy parts of making software that works. His growth lead Elena Verna does the same from the strategy side. Their chief designer Felix Haas shares the craft. It resonates because their audience — developers, tech leads, founders — can tell the difference between people who are in the work and people who’ve been handed talking points. In a market drowning in AI hype, that credibility is worth more than any ad buy.

Gong built a substantial chunk of their B2B brand the same way. Not just their executives on LinkedIn, but their sales team, their product people, their individual contributors — all sharing what they’re learning, what’s working, what’s not. It creates a network effect that a single corporate account, no matter how well-managed, could never replicate. You’re not following a logo. You’re following a dozen people whose thinking you’ve come to trust.

And Shopify and Figma and Framer have been doing this for years with their developer and design communities — empowering the people closest to the work to share what they’re building. Not because a marketing brief told them to, but because the culture actually supports it.

Now, the risks. Because there are real ones, and pretending otherwise is its own kind of corporate dishonesty.

People leave. They take their audiences with them. Your star product manager who’s been killing it on LinkedIn for two years? She gets poached by a competitor, and now she’s killing it for them. That’s a real cost.

But here’s the thing: she was going to build that audience whether you helped her or not. The question is whether you benefit while she’s with you, or whether you just watch from the sidelines, nervous about “investing in someone who might leave.” That calculus never works out. The best people always have options. The goal is to make staying more attractive than leaving — and supporting someone’s professional growth is part of that equation.

There’s also the risk of someone going off-script. Saying something that doesn’t align with brand values, or wading into territory that makes legal nervous. This is real, but it’s manageable. You need guidelines — not scripts, guidelines. Here’s what’s fair game, here’s what needs review, here’s who to call if you’re not sure. And the review process has to be fast and human. The moment it becomes bureaucratic, people will either stop participating or stop asking permission.

The other concern I hear is that this takes time away from “real work.” But the best employee content isn’t separate from the work — it is the work, narrated. Someone sharing the technical challenge they just solved. A customer win they’re genuinely proud of. A design decision that took three iterations to get right. You’re not asking people to become content creators on top of their jobs. You’re asking them to occasionally share what they’re already doing with a slightly wider audience.

According to Edelman’s Trust Barometer from as long ago as 2019, employees are now among the most credible spokespeople for a company — more trusted than CEOs, more trusted than PR representatives, more trusted than advertising. That’s not a trend. That’s a structural shift in how people evaluate who and what to believe.

The brands that figure out how to work with this — authentically, not performatively — will have an edge that compounds. The ones that ignore it, or try to manufacture it with mandatory share programs and canned content, will keep wondering why their organic reach is dying and their paid costs keep climbing.

But I want to be clear about the prerequisite: you can’t fake this. If your culture is broken, employee advocacy will surface that brokenness. If people don’t actually like working at your company, asking them to talk about it publicly is just handing them a microphone to share their ambivalence.

The organizations winning here are the ones where people genuinely want to talk about their work. Not because marketing asked them to, but because they’re proud of what they’re building. Because they feel like they’re part of something worth mentioning.

That’s not a content strategy. That’s a culture. And the content just follows.

Leave a comment

*This article was originally published on rebeccaraebarton.com