Your Next Customer Is an AI. Here’s What It Actually Reads.

The buyer reading your product page is increasingly not a person.

Ask ChatGPT for a pomodoro timer and it hands back products — images, prices, the details — inside the chat, without you opening a single retailer’s website. That’s the opening of Kartik Hosanagar’s June 2026 piece in Harvard Business Review, “How Do You Market to an AI Customer?”, and it names the shift plainly: an agent now sits between your product and the human, and it controls what gets discovered and what gets shortlisted. The human still decides. But more and more, the human only sees what the agent already chose to show.

So here’s the answer this piece hands over, up top: an AI customer skips the half of marketing built to persuade humans — the feel, the brand mood, the beautiful page — and weighs the half it can parse and check: structured facts, machine-readable specs, and claims it can verify. Your job flips from making a person feel good to making your product legible and provable. Evidence over vibes.

This is for you if your work is getting a product chosen — marketer, founder, product or content lead. If you’ve never thought about who (or what) reads your page before a human ever does, this is the part worth reading twice.

The agent is already the buyer

This isn’t a forecast about 2030. The plumbing shipped.

On September 29, 2025, OpenAI and Stripe published the Agentic Commerce Protocol (ACP) — an open standard for letting an AI agent discover products, build a cart, and complete a purchase on a person’s behalf. It went live the same day inside ChatGPT as “Instant Checkout,” starting with Etsy and rolling out to over a million Shopify merchants including Glossier, Vuori, and Skims. The spec is public on GitHub, maintained jointly by OpenAI and Stripe, and it has three parts: a product feed (so the agent can find and compare your catalog), a checkout API, and delegated payment. Amazon’s Rufus, Perplexity’s shopping, and Google’s Gemini are pushing the same pattern from their own ends.

How big this gets is still a forecast, and forecasts vary — Bain estimates US agentic commerce could reach $300–500 billion by 2030 (15–25% of e-commerce); Morgan Stanley pegs it at 10–20% of US e-commerce by the same year. Treat the exact number as a guess. The direction isn’t one. A real layer of buyers now reads your product before any human does, and Hosanagar’s warning is the sharp end of it: if the agent becomes the gatekeeper, the brands that don’t show up well to the agent risk becoming back-end fulfillment for someone else’s recommendation.

The half of marketing that just went quiet

Marketing was built to move a human: emotion, a brand you trust on feel, social proof, a hero image that makes the thing look like the life you want. An agent doesn’t have that wiring. It doesn’t get a warm feeling from your typeface. It doesn’t read your founder story and tear up. Show it an adjective like “premium” and it has no way to price that in — it’s a word, not a fact.

What an agent does do is read structured data, line up your product against the others against the buyer’s stated criteria, and check the claims it can check. So the persuasion layer — the part that worked by making a person feel something — goes quiet when the reader is a model. The half that survives is the legibility layer: can the agent find you, parse you, and confirm what you said is true?

That splits the work into three concrete moves.

1. Being found: from ranking on a page to being legible to an agent

Old discoverability was SEO for humans — rank on the page, win the click, the person scrolls and chooses. An agent doesn’t scroll a results page and admire it. It goes where the structured data is and parses it.

That means being findable and parseable in the places the agent actually looks — a clean product feed, structured markup on your pages, an API the agent can query — not just a page that ranks. A page tuned only for a human reader can be effectively invisible to a model that’s reading catalogs, not browsing. The first question is no longer “where do we rank?” It’s “can a machine read our catalog at all?”

2. Winning: from vibes to evidence

Once the agent has found you and your competitors, it compares. And it compares on what it can verify.

Specific, checkable numbers beat adjectives the model discounts. “Holds 1.2 liters, ships in 2 days, 4.6 stars across 3,400 reviews” is something an agent can line up against three rivals. “Premium hydration experience” is noise it skips. Third-party signals it can confirm — review counts, certifications, independent test results, return rates — carry more weight than anything you say about yourself, because the agent can cross-check them. This is the same instinct good buyers always had, made literal: a model can’t be charmed, so it leans on what’s provable.

The practical shift: every claim that decides a sale should be a claim a machine can verify. Swap the adjective for the number. Put the proof where it can be read, not just felt.

3. The shelf: from a landing page to a feed

Here’s the one that reorganizes the team. The unit of marketing stops being a hero image and becomes machine-readable product data.

This already has real specifications. Schema.org’s Product type lets you state price, currency, availability, and condition as structured fields a machine reads directly — Google’s own docs note it can pull prices, availability, brand, and ratings “without parsing visible HTML.” Google is moving to unify that Schema.org markup with its Merchant Center feed so there’s one source of truth. The ACP product feed asks for the same backbone — title, description, image, price, availability — as the catalog an agent indexes to present and compare you.

The work here isn’t art direction. It’s data hygiene: accurate, complete, current fields, in the format the agent reads. A gorgeous landing page with a stale or missing feed is a storefront with the lights on and the door locked.

This isn’t “brand is dead”

It’s tempting to read all this as the end of brand. It isn’t.

Humans still set the criteria the agent shops against — “find me a durable, ethical, under-$40 one” carries everything brand has always carried. Trust still earns you the shortlist: an agent leans on signals of reputation, and a person is far likelier to greenlight a name they already trust when the agent surfaces it. What changed is the price of persuasion, not its existence. The expensive, emotional, top-of-funnel work that used to be the whole game now gets you onto a shortlist that a machine, not a mood, finalizes. Persuasion gets cheap. It doesn’t get gone.

If that pattern feels familiar, it’s because you may already be living it at work. When AI took over the typing half of a lot of jobs, typing got cheap and judgment got paid — the skill didn’t vanish, its price moved. The same move is now hitting how you sell. The persuasion you used to spend the whole budget on gets commoditized; the legible, provable, machine-readable substance underneath it becomes the thing that wins.

Our work changed when AI started doing it. Our customers change when AI starts buying. Same shift, new surface.

The next move is a test, and you can run it today: could an agent find your product, read its specs, compare it against a rival, and verify your top claim — using only what’s machine-readable right now? Hand ChatGPT or Gemini your product and a competitor’s and watch what it can and can’t see. Whatever it can’t read is the part of your marketing that, to your next customer, doesn’t exist yet.