Your next customer is a bot

The next customer who considers your product may never see your homepage. Not the hero video, not the founder story, not the photography you argued over for a month. It arrives holding a goal (trail-running shoes, under $150, delivered by Friday), parses that sentence into constraints, queries structured data, ranks what comes back, and buys. Everything you built to persuade a human plays to an empty room. The buyer in this transaction is an AI agent, and it reads fields.

This is a field synthesis, not a prophecy. We read the specification OpenAI published for its own shopping surfaces, the first research that measures what machine buyers actually reward, and two years of Adobe's traffic data on the channel. The through-line is uncomfortable and simple: at the moment of machine decision, the brand that cannot be read as data is not persuading badly. It is not present at all.

Section OneThe funnel collapses into a query

The marketing funnel assumes a journey. Awareness, consideration, the emotional click, the homepage visit, the cart. Every stage is a surface where persuasion can work on a human: a color, a headline, a testimonial placed just so. The entire discipline of conversion optimization is the craft of nudging a wandering person toward a decision.

An agent does not wander. It interprets. "Trail shoes, under $150, by Friday" is not a mood; it is three constraints: a category, a price ceiling, a delivery deadline. The agent expands the goal, queries the product data it can reach, filters on the constraints, weighs the candidates, and returns one answer or three. There is no dwell time to measure, no scroll depth, no retargeting pixel to fire. The "visit," if one happens at all, is a structured request that lasts milliseconds.

The funnel doesn't shrink. It collapses from a journey into a query.

What survives the collapse is whatever can serve as an answer to that query. A price is an answer. An availability flag is an answer. A 4.7 rating over 2,000 reviews is an answer. A brand film is not an answer to anything a constraint can ask. This is not an argument that emotion stopped mattering to humans; it is an observation that a growing share of purchase decisions now pass through a reader that has no slot for it.

Consider what that does to the craft. A decade of conversion work optimized the seconds a human spends deciding: the button color, the urgency banner, the exit-intent popup. None of those surfaces exist in an agent's transaction. The optimization surface moves from the pixel to the record: is the price current, is the availability true, is the delivery promise machine-checkable, does the product's description say what the product is in words a parser can map to a constraint. Unglamorous questions. They are now the storefront.

Section TwoThe spec is public. Read it.

You do not have to speculate about what a shopping agent reads, because OpenAI published it. The Agentic Commerce product feed specification defines the exact data ChatGPT's shopping experiences consume. In the document's own words, it "defines the shared flat-file schema that OpenAI ingests and indexes." Field names, data types, constraints, down to example values.

OpenAI, "Product Feed Spec," Agentic Commerce developer documentation (live spec, checked August 2026).

The required substance is a merchant's least romantic data: item ID, title, description, price, availability, image URLs, brand, plus two flags, is_eligible_search and is_eligible_checkout, that decide whether your products can be surfaced and bought inside the answer at all. Whether the largest AI shopping channel can see your catalog is, literally, a boolean you set.

Two things about this document deserve emphasis. First, it is OpenAI's own feed format (structured product data, not Schema.org markup), and other agent surfaces will define their own interfaces. The specifics will vary; the shape will not. The interface to the machine buyer is fields. Second, notice what is absent. There is no field for your story. No field for the design award. No field for how the shoe makes anyone feel. The schema is the consideration set's front door, and it admits data.

Section ThreeWhat the machine buyer rewards

If agents make the selection, the next question is what moves them. The first serious measurement we have is a 2025 preprint by Allouah, Besbes, Figueroa, Kanoria and Kumar, who put AI agents through simulated purchase decisions and measured what changed the outcome. One preprint, not yet peer-reviewed, a single research group. Treat it as an early reading of the instruments, not settled science. Its direction, though, is hard to ignore.

Allouah, Besbes, Figueroa, Kanoria & Kumar, "What Is Your AI Agent Buying?", arXiv:2508.02630, 2025 (preprint).

Ratings weigh in far more heavily than they do for human shoppers, heavily enough to be priced. In the authors' words: "An increment of +0.1 in ratings, allows a seller to increase their price by a quarter with Gemini 2.5 Flash and a third with Claude Sonnet 4 while this more than doubles to 67% with GPT-4.1." One tenth of one star bought between 25 and 67 percent of price headroom, and the exact figure depends entirely on which model is doing the buying.

One tenth of a star buys a quarter of your price. Which machine is shopping decides how much more.

That model-dependence is itself a finding. Your product does not have one machine-facing reputation; it has one per agent, and they disagree. What is stable across all three models is the mechanism: the agent leans on the structured trust signals it can read (ratings, review counts, price), and leans on them disproportionately. The inputs to that math are exactly the fields from Section Two.

Follow the money one step further and the operational implication lands. If a tenth of a star is worth a quarter of your price to the buying machine, then review hygiene stops being a customer-service chore and becomes a pricing asset. The unanswered complaint, the unclaimed listing, the review request you never send: in a human market those cost you reputation slowly. In an agent-mediated market they compound into a number that sits directly in the selection math, next to your price, every single time the query runs.

Section FourThe receipts, dated honestly

Is any of this actually happening at volume? The traffic side is unambiguous. Adobe Analytics reported in March 2025 that traffic to US retail websites from generative-AI sources had grown 1,200 percent, February 2025 measured against July 2024. A small base, but that is what the beginning of every channel that later mattered has looked like.

Adobe, "Adobe Analytics: Traffic to US retail websites from generative AI sources jumps 1,200 percent," March 17, 2025.

The conversion side demands more honesty, because the metric has been volatile, and the volatility is the story. In July 2024, Adobe measured AI-referred visitors converting 43 percent worse than other traffic. By February 2025 the gap had narrowed to 9 percent worse. Adobe's later reporting then flipped the sign entirely: 42 percent better in its Q1 2026 reporting, 54 percent higher in data through May 2026, both as carried by trade press from Adobe's own panel. Four figures, four measurement windows, one vendor's self-measured panel, and a metric that swung from deeply negative to strongly positive in under two years.

Put the three exhibits together. The traffic is growing at four digits. The visitors it delivers now appear to convert at least as well as anyone else. And the selection step in the middle is increasingly performed by software that reads a published schema. That is not a future scenario. It is a distribution channel with a spec sheet.

In ClosingAbsent, not outranked

Classic search was forgiving in one specific way: you could lose visibly. Page two of the results was purgatory, but you existed there, and you could climb. An agent's answer has no page two. If your inventory, prices, availability and ratings are not structured where the machine reads, you are not ranked lower. You are not in the consideration set at all. The homepage stays beautiful. The buyer never enters the room.

No structured signal, no consideration set. Not penalized: absent.

The work this implies is not mysterious, and most of it is not even new. Publish your catalog as structured data where the agents read, and keep the boring fields (price, availability, delivery) ruthlessly accurate, because a machine trusts a stale field exactly once. Write product descriptions that state what the thing is before how it feels. Treat ratings as the pricing asset the early evidence says they are. None of this replaces the brand work that makes humans care; it is the translation layer that lets machines act on it.

This is the claim underneath everything we publish, and it keeps proving itself in new arenas: machines can only act on what they can clearly read: the buying agents outside your company, and increasingly the working agents inside it. Legibility used to be polish. In a market where the customer is a machine, it is the entry ticket, and the sellers who treat it as survival rather than styling will quietly absorb the demand of everyone who didn't.

If you want to know how visible your brand already is to the machines, you can measure it at /signal-index/. Or write to us at /contact/.