How AI shopping agents pick products is measurable, and the choice comes apart into four separate things that steer it. Where the product sits in the list. Which model is doing the choosing. What signals the platform attaches to the listing. How structured the product's own data is. Each behaves like its own axis, and each has been measured.
Most of that measurement comes from What Is Your AI Agent Buying?, a 2025 study by academic and industry researchers who audited shopping agents in a controlled marketplace sandbox across models from Anthropic, OpenAI, and Google, and their later versions. Each axis below carries one of that study's findings, and the fourth is corroborated by a second paper that reaches the same conclusion from a different method.
Position: does slot order change the pick?
Where a product sits in the list changes how often an agent picks it, and the effect is model-specific. The direction can reverse from one model to the next. The bias persists in headless, text-only settings with no visual layout, so it comes from how the model reads a sequence.
Model: does the answer change by which one is choosing?
Given the same shortlist, different models select at different rates, and updating one of them reshuffles the ranking. In one category the same fitness tracker was picked 45% of the time on one Claude version and 77% on its successor, while its share on OpenAI's models fell from about 25% to 6% across an update. Same product, same listing, opposite swings.
Signals: do platform tags shift the pick?
Sponsored tags tend to be penalized. Endorsement-style tags like "overall pick" tend to be rewarded. The direction mirrors how people treat ads and recommendations, and it holds consistently across the tested models.
Data: does clearer product data change the pick?
Clearer, more complete, machine-readable product data makes an agent more likely to pick a listing, and it can beat a higher rating. In the study, a seller shifted its own share upward with simple, query-matched description changes.
A separate 2026 machine-learning study by Bence Török at Eötvös Loránd University reaches the same result from a different method, modeling agent choice across thousands of simulated tasks and finding structured data the strongest single predictor of what gets picked, ahead of price and rating. Two studies, two methods, same lever. The Török study is covered in more depth on The Desk.
How to expose data this way... schema.org markup, server-side rendering, stable identifiers, semantic HTML... is on the Readiness guide.
Test it yourself
The headline finding is easy to reproduce. Take one of your products and three or four competitors, paste the same shortlist into ChatGPT, Claude, and Gemini, and ask each one to pick the best option for a specific shopper. Run it more than once. The choice changes between models, and sometimes between runs of the same model, which is the model-specific instability the research describes, visible in a few minutes with no tooling.
What's durable, what's in flux
Durable. Two things hold across models. First, that agent choice is measurable at all. The pick-test above works on any set of products and any set of models, and it will still work as models change. Second, clearer product data is the axis with the most consistent direction of effect across the tested models. The mechanics live on the Readiness guide.
In flux. Everything else moves. Position bias appears on every model measured, but its direction and magnitude change per model and change again on model updates. The specific per-model preferences (which product a given model tends to pick) are the least stable thing on this page. Platform signals (sponsored penalized, endorsement rewarded) have moved consistently so far, though the sample of tested surfaces is small.
Further reading
Where I have gone deeper on these on The Desk:
- LLMs are Better Sellers than Buyers ... a negotiation study and what a systematic buyer disadvantage means for buying agents.
- Good SEO Is a Head Start With AI Shopping Agents ... how structured data shapes which product an agent picks, and whether it can finish the purchase.