Last year I spent most of my time on one question. When an AI agent shops, what actually decides what it buys, and based on this answer, is there any way to apply it in the real world?
An agent lands on a store. Eight roughly similar products. It has to pick one and check out. What tips it?
I came at this from the research side, at My Custom AI, before I moved to the infrastructure side at Rye. The team's paper there, What Is Your AI Agent Buying?, was about exactly this... how an agent chooses, and whether it chooses the way a person does. Two more papers have landed on the question since. And they point at an answer I did not expect to be this... simple? Though the longer I sit with it, the more it is the one that makes sense.
The thing that most reliably gets a product chosen, and gets the purchase finished, is clean, structured product data. The same SEO hygiene you already owe your own website.
Let me walk through how I landed there.
The agent is its own kind of buyer
A shopping agent does not behave like a person, and it does not hold still. It carries biases a human does not, and those biases move from one model to the next. The My Custom AI work found position effects strong enough, and specific enough to each model, that the favored slot on one model is the least favored on another, and a single model update can flip it. Swap the model behind the agent and the products it prefers can reshuffle overnight.
So a lot of the levers people reach for are quicksand. Tune your listing to the position bias of today's model and the next release undoes the work. You would be optimizing against a target that moves every few months, and moves in a different direction for every provider.
Which raises the question I am most interested in... is there anything about a merchant's listing that an agent rewards no matter which model is driving it?
And there is... how legible your product is.
When the information an agent needs is structured, complete, and easy to parse, every model can use it, whether it is reading your web page or a plain feed of data. That is the part that holds still. Two newer studies put numbers on how much it matters.
What decides the pick, and the purchase
The first, a 2026 machine-learning study of autonomous shopping-agent choice by Bence Török at Eötvös Loránd University, modeled which product an agent selects out of a set, weighing price, ratings, reviews, brand, availability, and machine-legibility signals like structured data and description completeness. It comes at the My Custom AI question with a different instrument. Where that work ran live agents through a randomized storefront, this one models the choice statistically, across thousands of simulated shopping tasks. Same neighborhood, different method. The strongest single predictor of what got picked was structured data. Stronger than relative price. Stronger than average rating. A product with clearer structured information could win over a higher-rated rival, and good structure took some of the sting out of a higher price. When the weak listings in the data were cleaned up, the predicted winner changed in about one task in seven... and every new winner was a listing that had been improved.
I hold those exact numbers loosely, and I want to be straight about why. The dataset is synthetic, built to model agent choice rather than logged from live agents on live stores. So I read it for the direction it points, and leave the exact figures alone. The direction is clear, and it agrees with everything else.
The second study, Designing Agent-Ready Websites for AI Web Agents, is more concrete. Real browser agents, two versions of the same store, same products, same prices, same checkout. One version exposed its data as structured data, labeled its buttons and fields, and kept its details current. Three agents, three hundred runs. The clean version finished the task 89% of the time. The other managed 49%, with more wrong turns and higher cost getting there.
So the same structured data shows up in both studies, doing two different jobs. In Török's, it tips the decision... which product the agent picks. In Designing Agent-Ready Websites, it drives the execution... whether the agent can get through checkout once it has picked. The pick and the finish, and structured data sitting under both.
One thing sits underneath all of it, though, and merchants run into this one first. An agent only reads your structured data or clicks your labeled button if it can reach your site at all. Plenty of stores still turn agent traffic away at the door, through the same bot detection built to stop scrapers. Legibility does nothing for an agent your firewall already blocked. Whether to let agents in, and how to tell a shopping agent from a bad bot, is its own problem and its own piece. This one assumes the door is open.
None of this is new work
Notice what none of this asks you to build. Structured data is schema.org, the markup that has earned rich results in Google for a decade. Labeled buttons and semantic HTML are what accessibility has asked for since screen readers existed. Complete, current product copy is on-page SEO by its oldest definition.
It is worth separating this from the agent-only rails people keep proposing. I wrote recently about WebMCP, where a site hand-writes a set of tools just for agents and keeps them in sync as the site changes. A separate surface, a real maintenance load. Neither study here tested anything like that. Both leaned on plain markup on an ordinary store. So you get the lift from work you already had reason to do, without standing up a new interface for agents at all.
And you can start in-house. This is your existing web team's work... structured data, semantic markup, honest product copy... not a vendor you have to hire before you begin. The teams building deeper agentic commerce optimization are doing real work, and this hygiene is the ground it stands on. The ground is yours to lay today, on the site you already run, and the same effort pays off for the human who lands on the page.
What I would actually do
If I ran a store and wanted one move this quarter, it would be the structured data. Put the product facts an agent needs into schema.org/Product markup... price, availability, variants, specs... so a model reads them the same way whether it parses the page or pulls a feed. That is the signal the selection study leaned on hardest.
Then the parts that let an agent take action. Label your buttons and fields with plain names, so an agent can act on them without guessing. Give the interactive pieces stable, explicit identifiers, so the path through checkout is unambiguous. This is where the 89-versus-49 gap lived... fewer wrong turns, fewer dead ends.
Then keep your evidence on the page and current. Specs, reviews, shipping and return terms, with dates on them. An agent leans on that to justify a choice, the way a careful shopper does. And write your descriptions in plain, specific language, because the clearer the content and context, the more reliably an agent reads what you sell and who it is for.
None of that is exotic. It is ordinary web work, and where your site starts decides how much of it there is to do.
Where this stops
Two honest limits. These are early studies, one synthetic and one a sandbox test, so read them as strong signals and not settled law. And legibility is not a stand-in for a good product. The research is careful here, and so am I... clean structure helps an agent understand what you offer, it does not turn a weak product into a good one. Török raises exactly this. Better documented is not the same as better, so a brand with the resources to run clean data can edge out a smaller one selling a better product, and firms can end up polishing the listing instead of fixing what they sell.
So if you want one thing to do about shopping agents, do the boring, durable thing. Fix your structured data. Label your page. Write your product like you want it read, because more and more, something is reading all of it. The models will keep changing under you. Clean, legible product data is the part that does not.
Frequently Asked Questions
How do you make a website ready for AI shopping agents?
Focus on machine legibility, which is the SEO and accessibility hygiene you already owe your site. Expose product data as structured data with schema.org/Product, use semantic HTML with clearly labeled buttons and fields, keep product evidence like specs, reviews and policies current, and write descriptions in plain, specific language. Research on agent behavior shows structured data is both the strongest signal in which product an agent picks and a major factor in whether it can complete checkout.
Does structured data really affect which product an AI agent chooses?
In a 2026 machine-learning study of autonomous shopping-agent choice, structured data was the strongest single predictor of what an agent selected, ahead of price and average rating. A clearer, better-structured listing could be chosen over a higher-rated rival. The dataset is synthetic, so treat the exact figures as directional, but the direction lines up with other agent-behavior research.
Do I need a special agentic version of my website or a new protocol?
No. The gains in the research came from plain markup on an ordinary store, structured data, semantic HTML and current evidence, not from a separate agent-only interface like WebMCP or a new commerce protocol. It is your existing web team's work, on the site you already run, and it also helps human shoppers.
What stops an AI agent from buying before structured data even matters?
Access. An agent can only read your structured data or use your checkout if it can reach your site at all. Many stores block agent traffic with bot detection built to stop scrapers. Deciding whether to allow shopping agents, and how to tell them from bad bots, comes before any of the legibility work.