Make Your Brand Machine-Readable

(Was this newsletter forwarded to you? Sign up here.)

Last week I argued that the agent is going to do a lot of the buying, and that when it does, you lose the data, the upsell, and the relationship even when you keep the sale. A few of you wrote back with a fair question: fine, but is there anything I can actually do about it, or do I just watch?

There's plenty to do. It just looks nothing like what you're used to.

Start with the scale of it. Visa, launching its agent-verification protocol last fall, cited a 4,700% surge in AI-driven traffic to US retail sites. Your next buyer is increasingly a piece of software doing research on a human's behalf, and software cannot be charmed, retargeted, or made to feel anything. Everything in your playbook assumes a person on the other end. The person is leaving.

Everything you do to a shopper backfires on a machine

We have hard evidence on what actually moves an agent, and it's one of the most useful marketing study I've read this year. Researchers built a shopping-agent simulator and ran the leading models, Claude, GPT, and Gemini, through thousands of purchase decisions while manipulating one variable at a time. The results should rearrange your priorities.

Tag a product as sponsored, and the agents bought it less. A product that won 10% of decisions dropped to around 8% the moment it was labeled advertising. The machine discounts paid placement on purpose. Add a scarcity or countdown-timer tag, the move that reliably lifts human conversion, and the effect was weakly negative or nothing at all. The agent does not feel urgency because it does not feel.

Now the inverse. Give the same product an editorial endorsement, an "overall pick" style badge, and its share jumped from 10% to 24% on some models, and as high as 42% on others. That badge is worth so much a brand could raise its price by more than half and still get picked. Raise its star rating by a tenth of a point and share climbed to 15 to 20%. Double its review count and the agent tolerated a 17 to 37% higher price.

Read that again. The things you can buy your way into, ads and placement and urgency, move the agent down or not at all. The things you have to earn, ratings, reviews, and third-party endorsement, are exactly what it rewards. The human funnel runs on persuasion. The machine runs on evidence, and it is brutally rational about which evidence counts.

The Three Things a Machine Reads

So stop thinking about messaging and start thinking about what an agent can actually parse. There are three things it reads, in this order, and they map to a simple funnel: can it see you, will it pick you, will it name you.

The first is your feed. Before an agent can choose you, you have to be machine-visible, and most brands aren't. OpenAI's commerce feed has a field called is_eligible_search, and it defaults to false. If no one flipped it, you are invisible inside ChatGPT. Not ranked low. Absent. The same feed wants a stable product ID that never changes, a product URL that returns a clean 200, and price and availability that are actually accurate. Google's version is a public manifest you host at /.well-known/ucp that tells agents what they can do with you. This is plumbing; it is unglamorous, and it is the difference between being a candidate and being absent.

The second is your reviews. This is where selection happens, and it's where the study above lands hardest. Ratings and review volume aren't social proof for a human reader anymore; they're the primary input an agent weighs when it ranks the options it found. Review count literally buys price headroom in the model's decision. If your reviews are thin, stale, or stranded on a property the agent doesn't read, you lose the comparison before a human ever sees it, and you never find out why.

The third is your reputation, by which I mean how often the rest of the web mentions you. When Ahrefs correlated AI visibility across 75,000 brands, the strongest signal wasn't backlinks, which barely registered. It was branded mentions across the web, with YouTube mentions the single strongest factor they measured. The agent decides who makes the consideration set based partly on who the internet talks about, and that conversation happens mostly off your own site, on Reddit, YouTube, Wikipedia, and comparison content. Each engine even has a different diet: ChatGPT leans on Wikipedia, Perplexity on Reddit, Google's overviews somewhere in between.

But isn't this just GEO snake oil?

Reasonable suspicion. Every agency on earth bolted "generative engine optimization" onto its deck this year, and most of it is repackaged SEO with a markup. 

First, doesn't this all change with the next model? Partly yes. The study found that an agent's position bias, which slot it favors, flipped completely between two versions of the same model, and you cannot durably game that. But the durable levers, clean feeds, real reviews, and genuine brand mentions aren't model-specific tricks. They're the inputs every model reads because they're the inputs that describe reality, and reality doesn't get patched.

Second, can you even measure it? Better than you'd think, and worse than the vendors claim. The monitoring tools that watch your presence across engines are real and worth running. The "we boosted AI traffic 186%" case studies are mostly marketing. Buy the measurement, not the promises. And skip llms.txt, the file the GEO crowd keeps selling. No major engine has confirmed it uses it.

Your checklist

Hand this to whoever owns your site and feeds:

• Confirm the product feed is live and is_eligible_search is set to true.

• Verify product URLs return a clean 200 and product IDs are stable.

• Publish the /.well-known/ucp manifest.

• Audit your WAF and robots rules so you aren't silently blocking the AI search crawlers (OAI-SearchBot, PerplexityBot, and their kin). A blocked crawler is a brand the agent can't cite.

Then point real budget at the two earned levers: a relentless engine for collecting recent third-party reviews, and a brand-mention program (digital PR, podcasts, YouTube, comparison content) that gets the rest of the web saying your name. Stop spending to sway the agent with placement; it's built to discount.

None of this is the work you were trained to do. It's closer to plumbing and PR than to campaigns. But the buyer changed, and the buyer can't be persuaded, only described accurately enough, and reviewed well enough, and mentioned often enough, that the math comes out in your favor. The brands that figure this out won't be the loudest. They'll be the most legible.