Can what you say you want, without ever naming a price, reveal exactly what you'd pay?

This explores whether a buyer's described tastes and priorities, with no budget number anywhere, can tell an observer what that buyer is willing to pay. The corpus has a strong, somewhat surprising answer: yes, and in one experiment the descriptive version worked better than the number. When an LLM buyer acted out a verbal shopping persona (who they are, what they care about, how they shop), a seller-side agent reading only the dialogue recovered the buyer's target price almost perfectly, with a slope of 1.00. When the buyer was instead given a confidential numeric budget, the seller's inference dropped to a slope of 0.21 Can a seller infer what a buyer really wants to pay?. A consistent character leaks its price point through dozens of small choices. A hidden number only shows up if it's said out loud.

The twist is that numbers handled by AI agents tend to come out anyway, and badly. In negotiation tests across GPT, Qwen and DeepSeek models, buyer agents disclosed their budgets for no reason, accepted inflated prices and ignored their own cost limits. The researchers traced this to instruction-following failures, not strategy Do LLM negotiation agents reliably follow budget and pricing rules?. Agents also overreact to the price cues they do see. Compared with humans facing identical product choices, LLM shoppers swung far harder in response to price, rating and nudge changes, falling into rule-like patterns where humans made modest, probabilistic shifts Do shopping agents respond more strongly to price and rating cues?. Put the three findings together: a stated budget is a weak and fragile signal for agents, while a coherent description of preferences is a rich one, which is good for sellers and possibly bad for buyers' privacy.

The personalization literature points the same way from a different angle. Compact, abstract preference summaries beat retrieving specific past interactions Does abstract preference knowledge outperform specific interaction recall?, and about ten well-chosen questions can be enough to pin down a person's individual preference weights Can user preferences be learned from just ten questions?. Some systems skip asking altogether and infer preferences by watching behavior over time Can agents learn preferences by watching rather than asking?. If preferences compress this well, it makes sense that a price point would fall out of them.

There is a catch, though: 'stated preferences' aren't one thing. Behavioral science work on annotation data separates genuine preferences from non-attitudes (answers given without any real opinion) and from preferences constructed on the spot by the way the question was asked. Only the first kind holds steady across conditions Do all annotation responses measure the same underlying thing?. The persona experiment worked partly because the simulated buyer's preferences were coherent by design, and real shoppers are messier. Preferences may not even sit still: one argument holds that AI shopping agents which execute purchases end up co-shaping what consumers want through repeated conversation loops Do AI shopping agents reshape what consumers actually want?. If that's right, an agent inferring your willingness to pay may also be helping set it.

The headline is that how you talk about what you want can reveal your price better than naming one. The open question the corpus can't yet settle is whether that holds for real humans, whose preferences are inconsistent, and not just for LLM-simulated personas.