If a shopper hides money talk from their profile, can sellers still figure out their real budget from everything else they say?

This explores whether stripping explicit money talk out of a consumer profile actually protects how much someone is willing to pay, or whether the signal leaks through anyway. The collection doesn't have a study that removes financial vocabulary and measures the drop. It does have a closely related result that points the other way from what most people would guess. In a seller-side experiment, an inference agent recovered a buyer's target willingness to pay almost perfectly, with a slope of 1.00, from the shopping dialogue alone, but only when the buyer was playing a verbal persona. When the buyer instead held a numeric budget and kept it confidential, the seller's inference collapsed to a slope of 0.21 Can a seller infer what a buyer really wants to pay?. The leak isn't in the dollar figures. It comes from coherent behavior: a persona that acts consistently tells a seller more than a hidden number does.

So removing financial vocabulary may weaken the signal far less than you'd expect, and could even matter less than how vivid and consistent the profile is. A similar pattern shows up in work on what makes personalization effective. Abstract summaries of someone's preferences beat retrieving their specific past interactions Does abstract preference knowledge outperform specific interaction recall?. The general shape of who someone is carries more usable information than any single explicit detail. A profile that says "values craftsmanship, shops at independent boutiques, hates fast fashion" never names a price, but it says a lot about one.

There's also a reason to doubt that a single, clean "willingness to pay" exists to hide. Behavioral science work on annotation data splits people's stated responses into genuine preferences, non-attitudes (answers given without a real opinion behind them), and preferences constructed on the spot by how the question was asked Do all annotation responses measure the same underlying thing?. Bilgihan and colleagues argue that when AI agents carry out purchases, preferences get built through back-and-forth between agent and consumer rather than existing fully formed beforehand Do AI shopping agents reshape what consumers actually want?. If willingness to pay is partly built during the conversation, scrubbing words from a static profile misses where much of the signal comes from.

The agent side adds a twist. Shopping agents react to price, rating, and nudge cues much more sharply than humans do, falling into rule-like patterns instead of the modest, probabilistic shifts people show Do shopping agents respond more strongly to price and rating cues?. And user profiles tend to pull models toward pleasing the user rather than giving balanced answers Does personalization make large language models worse at their jobs?. Together these suggest that an agent with a lifestyle-rich profile may act out that persona very legibly, which is exactly the condition that let the seller infer willingness to pay perfectly. Models also let their own values shape answers without saying so Do language models leak their own values into practical advice?, which is a reminder that inference without explicit vocabulary runs in both directions.

The takeaway: the corpus suggests that hiding the numbers isn't the same as hiding the budget. If you're designing privacy for consumer agents, the more promising protection may be a sparse, confidential numeric constraint rather than a rich verbal persona with the money words removed. That conclusion is drawn from one comparison, though, not from a direct experiment that removes the vocabulary.