The Risks and Rewards of User Expressiveness in Revenue-Driven Recommendation

Large-language-model recommendation interfaces let users communicate much richer preference information than traditional, coarser search. Richer expression can improve product fit, but it also reveals \textit{precisely} where a user is willing to compromise---information a revenue-driven platform can use to favor higher-revenue items. We study this tension in a feature-based model where a user sends soft preferences over product features and optional hard filters to a platform with a private catalog. The platform recommends the item maximizing per-purchase revenue times a purchase probability evaluated at the reported match. We show that truthful expressive reporting can be worse than coarser truthful reporting. We then give a condition under which a strategic expressive user can find a near-optimal item in polynomially many interactions by trying a one-dimensional family of reports that counteracts the platform's revenue bias by down-weighting high-revenue features. When revenue differences are too large to be offset by soft reports, expressive reporting can require exponentially many interactions. Similarly, optimal reporting under coarse interfaces can also incur exponential search cost.