Price tags and star ratings sway AI shopping bots far more than they sway people — so who's really the easier sell?

This explores how AI shopping agents weigh numbers like price and star ratings against the wording of a product listing, and whether models differ in how they do it. The corpus has no single study that tests price against description head-to-head inside the same model. Several studies come at the question from different sides, though, and together they give a clear picture. Both kinds of signal hit AI agents much harder than they hit people. Which one wins depends less on the model than on how easy it is for the agent to look at the details.

Start with the numbers. When 17 models faced the same binary product choices as humans, they shifted their picks far more than people did in response to price, rating, and nudge changes (Do shopping agents respond more strongly to price and rating cues?). Humans nudge their preferences a little. Agents tend to collapse into rule-like behavior, such as "cheaper wins" or "higher rating wins." A shopper who reacts this mechanically is predictable, and that makes it easy to exploit.

Text turns out to be just as powerful. In a marketplace simulation, a seller-side AI rewrote only its product descriptions, with no change to price, rating, or position. That alone raised its market share by roughly 4 to 15 percentage points with five of the six buyer models tested (Can AI seller agents shift buyer preferences through description edits?). One model held firm, which suggests that how easily a model is swayed by wording is partly a quirk of that model. The same research group found that position bias (favoring whichever item appears first or in a given slot) is also strongly model-specific. It survives text-only interfaces and explicit instructions to ignore it, and it reappears after model updates (Do AI shopping agents show persistent position bias in text interfaces?). So each model brings its own built-in leanings to the shelf before it has read a single price.

The most useful finding comes from a study of eight commercial models. When every product attribute was visible, they mostly resisted marketing tricks. When checking details was costly and the user's goal was vague ("get me a good one"), they fell back on price as a shortcut (When do pricing cues actually mislead AI shopping agents?). The failure was in gathering information, not in reasoning or arithmetic. Price is a cheap signal to read. A description has to be read, and attributes have to be looked up. When looking is expensive, the number wins by default. The same pattern shows up beyond shopping. Six LLMs gave the same strategic advice almost regardless of context, while simply reordering the options shifted results more than changing the industry did (Do LLMs consistently favor the same strategic choices regardless of context?). AI search engines cite brands largely by how famous they already are (Do AI search engines favor well-known brands over others?).

One more twist: agents don't handle price well even when it's front and center. In negotiations across GPT, Qwen, and DeepSeek, buyer agents revealed their budgets, overpaid, and broke their own cost limits. These were instruction-following failures, not clever strategy (Do LLM negotiation agents reliably follow budget and pricing rules?). And in the other direction, a seller agent could work out a buyer's willingness to pay almost perfectly from a natural-language shopping persona, but much less well from a confidential numeric budget (Can a seller infer what a buyer really wants to pay?). Taken together: price is the signal AI agents grab by default when information is hard to get, and description text is the lever sellers can quietly pull. Both affect agents more strongly than they affect people, which is why the design of agent-facing storefronts may matter more than which model you pick.