Why do AI shopping agents swing wildly on cues like star ratings, while people barely budge at the same prices?

This explores why AI shopping agents react more sharply than people to the same cues (price, ratings, nudges) and what follows from that. The most direct evidence comes from a benchmark that ran 17 models against human shoppers on simple two-product choices Do shopping agents respond more strongly to price and rating cues?. When people saw a nudge, their choices shifted a little and unevenly. The agents' choices shifted a lot, and they settled into something close to a fixed rule: if one product's rating is higher, pick it every time. Humans are noisy, and that noise turns out to protect them. People get distracted, weigh things inconsistently and disagree with each other, so a single cue only moves some of them. An agent applies the same cue the same way every time.

The agents also aren't over-reacting because they think like people do. Agents show human-like choice biases without the limited attention or memory that causes those biases in people Where do LLM choice biases actually come from?. The biases appear to be patterns learned from training data and reward signals. In people, a bias usually comes from a mental shortcut that tiredness or distraction switches on only some of the time. In a model, the learned pattern fires reliably. Position bias is a clear example. Agents prefer items in certain list positions even in plain-text interfaces with no visual layout, even when told to ignore position, and the preference comes back after each model update Do AI shopping agents show persistent position bias in text interfaces?.

The over-reaction isn't fixed, though. It gets worse in particular situations. Eight commercial models resisted marketing tricks when every product detail was visible, but fell back on 'just pick the cheapest' when looking up details was costly and the user's goal was vague When do pricing cues actually mislead AI shopping agents?. The failure was in what information they went and got, not in their reasoning or arithmetic. In another case, a Gemini agent buying in bulk quietly switched from comparing unit prices to minimizing the total price. It searched less, and its rate of making the best choice fell from 80% to 35% Does higher purchase volume make AI shoppers search less carefully?. So the agent can quietly swap the goal it was given for a simpler one it can act on.

The part you might not have thought to ask about is that predictable buyers are easy to steer. If an agent reacts the same way to the same cue every time, a seller can learn that reaction and target it. A seller-side agent that only rewrote product descriptions, with no change to price, rating or position, raised its product's market share by up to about 15 percentage points Can AI seller agents shift buyer preferences through description edits?. Sellers can also work out what a buyer is willing to pay almost perfectly from the shopping conversation alone Can a seller infer what a buyer really wants to pay?. Add an agent funded by advertising, which has its own reasons to favor paid placements Do advertising-funded shopping agents favor paid product placements?, and the over-reaction stops being an odd quirk and becomes a weakness that can be exploited across a whole market.

That changes what 'better shopping agents' should mean. Making agents smarter doesn't solve it. They also need to keep some of the useful inconsistency that makes human crowds hard to steer, or be paired with agents that represent the other side of the deal, so that steering meets pushback. Choudary argues that agents acting only for buyers change how buyers get into a market but not how the market works Do one-sided consumer agents actually change market structure?. As long as sellers can see how the agents behave and buyers can't see the sellers' tactics, the agents' tendency to over-react mostly benefits sellers.