When an AI shops for you, does it obsess over price and star ratings way more than you ever would?

This explores whether AI shopping agents react more strongly than people do to cues like price and star ratings. The short answer from the corpus is yes, and by a lot. In a benchmark that put 17 LLMs and human participants through the same two-product choices, a small change in price, rating, or a nudge moved the agents' choices far more than it moved the humans' Do shopping agents respond more strongly to price and rating cues?. People shifted a little and kept some randomness, as if weighing several things at once. The agents snapped into something close to a rule: the cheaper item, or the higher-rated one, nearly every time. So the gap isn't only about how much they respond. Agents turn soft preferences into hard thresholds.

Why would a model act more like a rulebook than a shopper? One answer is that these biases look human but come from somewhere else. People lean on shortcuts like ratings because their attention and memory are limited. LLMs have no such limits, yet they still show human-like choice biases, which points to patterns learned from training data and reward signals rather than a shared way of reasoning Where do LLM choice biases actually come from?. That changes how to read the over-response. The agent isn't overwhelmed. It has learned that 'higher rating wins' and applies the lesson with more confidence than any human would.

Price and rating aren't the only signals agents overweight. Other studies find the same pattern with cues that shouldn't matter at all. Agents favor products based on where they appear in a list, even in text-only interfaces and even when told to ignore position, and the specific bias changes with each model update Do AI shopping agents show persistent position bias in text interfaces?. In simulated marketplaces, every model tested showed a strong first-proposal bias, weighting how fast an offer arrived 10 to 30 times more than how good it was Do LLM agents achieve fair marketplace outcomes when given ideal search?. Over-response, then, is a general trait: agents latch onto whichever clear signal they have and give it too much weight.

The twist is that over-reacting to price doesn't make agents good bargainers. In negotiation tests across GPT, Qwen, and DeepSeek models, buyer agents revealed their budgets without being asked, accepted inflated prices, and broke their own spending limits, while seller agents accepted offers below wholesale cost. The researchers traced these failures to poor instruction-following rather than strategy Do LLM negotiation agents reliably follow budget and pricing rules?. A related finding: when a buyer agent plays a verbal persona, a seller agent can work out its willingness to pay almost perfectly from the conversation alone Can a seller infer what a buyer really wants to pay?. Agents can be very sensitive to a visible price tag and still easy to read and exploit on the other side of the table.

What you might not have expected to want to know: predictable, rule-like reactions are exactly what sellers and platforms can optimize against. If a half-star bump or a small discount reliably flips an agent's choice, merchants have every reason to engineer those cues. Advertising-funded agents add another pull toward paid placements Do advertising-funded shopping agents favor paid product placements?. And because agents that buy for you also help shape what you end up wanting Do AI shopping agents reshape what consumers actually want?, an agent's over-sensitivity to price and rating can, over time, become yours. The corpus has strong evidence that over-response exists. It says much less about fixing it, beyond noting that the bias is model-specific and comes back after each update.