After the mention comes the choice: AI shopping agents barely weigh their options
English translation of the German original of 28 October 2026.
An AI agent is asked to buy a fitness watch. In front of it are eight models, plus a review site, a forum and a magazine. In new experiments, which watch it takes depends less on the watches than on which source it reads first, in what form it arrives and what it thinks it knows about the user. For GEO, this means: the mention in the AI answer is the entry ticket. The choice is made afterwards.
Two controlled experiments with around 26,000 and 23,015 test runs in simulated shops. The effects vary strongly by model.
- →Being mentioned in an AI answer does not mean being chosen. Two new experiments examine the step after that: how AI agents pick one product out of several.
- →Agents do not average across several sources. Usually one source wins, and which one depends on the model. The order of the sources, the way they are delivered and a single sentence of user context also shift the choice.
- →The more strongly a sales cue such as “bestseller” steers the choice, the fewer alternatives the agent checks before deciding.
- →Real systems use exactly these levers: for ChatGPT, OpenAI documents stored user memories and model answers generated in advance as inputs to product selection, creates labels such as “Most popular” itself, and then ranks the retailers in a second selection.
- →A representative survey in Germany shows: most people want AI to compare for them, but prefer to pay themselves. The preselection still happens in the AI.
- →Method: lab experiments with simulated shops and a limited product range each. They show how susceptible agents are, not how often this happens in real shopping.
Mentioned is not chosen
Our stage model in “From the matching page to the brand mention” (in German) ends with the mention: a piece of content matches the query, the AI picks it up as a source, the brand is mentioned. The study behind it went no further. What happens after the mention remained open.
For shopping agents, this question can now be examined experimentally. Researchers build simulated shops, give an agent a shopping task and deliberately change individual conditions: which sources it sees beforehand, in what order, with which sales cues. Then they count which product it chooses. Two such experiments have been published in recent weeks.
One source wins
A team at the Wharton School had six current models from Anthropic, OpenAI and Google buy a fitness watch around 26,000 times. Each time, the agent saw a grid of eight watches; each condition was run 200 times. Without further context, every model had a stable favourite. Then the researchers changed what the agent got to see before buying:
“Choices were stable on the bare grid; the added layers made them unpredictable.” — Kumar, Meincke, Shapiro, Mollick, Puntoni and Mollick, Wharton Generative AI Labs
The authors draw a conclusion mainly for developers: anyone who tests shopping agents only at the actual checkout interface underestimates how easily they can be redirected. For brands, there is a second lesson: which source an agent treats as decisive is hard to predict from the outside and differs from model to model.
The stronger the cue, the shorter the search
Researchers at KAIST and Arizona State University ran 23,015 test runs with three models from two providers in a simulated shop. They varied nine classic sales cues, that is, signals actually meant for human shoppers. The cues worked, but with different strength depending on the cue and the model. In the most important model tested, a “bestseller” cue raised a product's selection rate from 11.5% to 70.0%.
That a language model trained on human texts responds to such signals is hardly surprising. The second finding is more remarkable: the more strongly a cue steered the choice, the fewer alternatives the agent looked at before committing. The authors call this a trade-off between diagnosticity and exploration. A strong signal does not only shift the choice; it also shortens the checking beforehand.
Limitation: this is a short paper for the ICIS 2026 conference. Only the abstract has been published so far, and it does not name the models tested.
The same levers, officially documented
Lab experiments leave open whether real systems work this way at all. For ChatGPT, OpenAI itself published an updated description on 28 October 2026. According to it, a product appears in the selection if ChatGPT considers it a match for the user's intent. More than the query feeds into that assessment:
In OpenAI's example, anyone who has once said they do not like clowns will no longer be shown clown costumes. How much such context is stored is shown by a study from the Max Planck Institute for Software Systems, Ruhr University Bochum and TU Delft, presented at the ACM Web Conference 2026: of 2,050 memories about 80 real users, ChatGPT created 96% itself, without an explicit request; 52% contained psychological assessments of the person. The context that co-determines product choice therefore arises largely in passing, and neither users nor brands steer it deliberately.
Limitation: the data were collected before February 2026; OpenAI has developed the memory system further since then.
Remarkably open is the reference to “model answers that ChatGPT generates before taking new search results into account”. So the model already brings an opinion before it searches. How strong this effect is, we described in “Found or remembered” (in German).
Sales cues also exist in the real system, except that ChatGPT sets them itself. “Budget-friendly” can mean that reviewers often praise value for money, not that the product is the cheapest. So the system does not only read such signals; it also creates them. And in the subsequent retailer selection, the price in the first answer is usually that of the first retailer, not necessarily the lowest. Stronger personalisation of this ranking has been announced.
Limitation: the documentation names inputs, but no weightings and no tests of selection quality. It shows which levers exist, not how strongly they work.
AI preselects, humans pay
These findings do not require agents to soon shop autonomously. A representative online survey of 500 consumers, conducted by the EHI Retail Institute for KPMG and published on 8 October, shows the division of labour. 46% have used AI often or very often for purchase decisions in the past twelve months. Many are happy to leave comparing and researching to an agent; when it comes to paying, willingness drops markedly, and most want to be able to intervene at any time.
Online survey of 500 consumers in Germany, conducted by the EHI Retail Institute for KPMG, October 2026.
What stands out is whom respondents trust: 64% consider general AI assistants more objective, only 7% the AI of a retailer or manufacturer. So preselection preferably happens in systems that brands do not control themselves. That buyers often still check AI recommendations themselves, we described in “After the AI recommendation, the purchase is still open” (in German). But only what remains after the preselection gets checked.
Providers are building exactly this division of labour. On 23 October, Anthropic introduced Claude Duet, a proactive agent that works across more than 6,000 integrations, including Shopify and Stripe. It buys via Link or Shop Pay, but only after the user confirms. Similar agents exist from Google, OpenAI and Meta. Duet is so far only available as a preview in the US.
Claim: “A bestseller cue wins AI agents over.”
Assessment: Not that simply. In the experiment, the cue was set artificially. In real retail, such cues are usually assigned by the platforms, not the brands. The effect also varied by cue and model. And a false bestseller cue would simply be misleading.
Claim: “AI agents decide based on objective product data.”
Assessment: No. On a bare product grid, the models did choose consistently. But as soon as sources, their order or user context were added, the choice shifted, sometimes almost completely, including away from the objectively best offer.
Claim: “This only becomes relevant once agents pay by themselves.”
Assessment: No. As soon as an AI narrows five products down to two, the decision has largely been made. Many consumers already want to give AI exactly this role of comparing and preselecting.
Claim: “The results apply to all product categories and agents.”
Assessment: That is open. One experiment used a single category, fitness watches, and screenshots instead of real websites. Of the second, only a short version has been published so far, without naming the models tested.
Visibility opens the door, the decision is made behind it
The experiments extend the chain from the stage model by one stage that is hardly measured so far:
Visibility remains the precondition: what is not mentioned cannot be chosen. But between mention and purchase, the agent selects which product it proposes, and that selection follows different rules. It weighs less than assumed. It often follows a single source, reacts to the form in which information arrives, and stops searching earlier if a signal is strong enough.
For brands, this does not mean feeding agents with sales cues. It means that the statements about a product must hold up in all sources an agent might read: on review sites, in forums, in comparisons and in their own product data. If it is unclear which source tips the balance for which model, consistency across all sources is the most robust strategy.
“GEO does not end with the mention. Being mentioned gets you into the race. The winner is the one whose story holds up in the sources the agent happens to be listening to.”
Dr. Michael Mohaupt, rethink.expert
Five checks for the selection stage
Know the sources agents read: what do review sites, forums and comparison portals say about your product, and do they say the same? Contradictions are the biggest risk, because one source alone often tips the balance.
Measure the choice, not just the mention: in test runs with fixed candidates, check how often an agent chooses your product, across several models and with repetitions.
Prove advantages, don't just claim them: concrete data, prices and ratings carry weight even when no sales cue helps. Cues such as “bestseller” are usually assigned by the platforms.
For retailers: keep an eye on the second selection. A product can be chosen and your shop can still end up at the back. Keep prices and availability in product data up to date and check which retailer appears first for purchase-related queries.
Test your own shopping assistants realistically: with upstream sources and user context, not just on the bare product grid.
AI agents and product choice
In experiments, they weigh less than expected. Without context, every model has a stable favourite. When sources are added, one usually wins, and which one depends on the model. The order of the sources, the way they are delivered and one sentence of user context can also shift the choice.
Yes, clearly in the experiment: a “bestseller” cue raised a product's selection rate in one model from 11.5% to 70%. But the effect varies by cue and model. And the stronger a cue works, the fewer alternatives the agent checks beforehand.
Rarely. Many consumers want to use AI for comparing, but pay themselves. For product choice, that is enough: as soon as an AI narrows the selection, the decision has largely been made.
Whether the sources agents read say the same about their product, and how often an agent chooses the product when it is on the shortlist, across several models and repetitions. A mention alone says little about that.
How often does an AI agent choose your product when it is on the shortlist? Get in touch to test it across several models and source situations.
Get in touch →