Fit Cards for Agentic Marketplace Search: Query-Conditioned Structured Metadata to Reduce Welfare Loss at Large Consideration Sets

Abstract

As AI agents increasingly conduct economic transactions on behalf of humans, marketplaces face a new challenge: agents evaluating large consideration sets cannot reliably extract fit signals from truncated natural language descriptions. We show that agent-side interventions---prompting and inference scaling---fail to address this information bottleneck, with both approaches yielding negative welfare. We propose \textbf{Fit Cards}, a platform-side intervention that replaces truncated descriptions with query-conditioned structured metadata computed from the platform's catalog, including item availability, amenity matches, and estimated prices. In experiments on a restaurant marketplace with 100 businesses, Fit Cards achieve 11.4$\times$ higher welfare than the status quo baseline (783.78 vs 68.51), improve contacted-fit rate from 7.8% to 73.6%, and reduce LLM calls by 28%. Our results demonstrate that discovery quality---not proposal-stage negotiation---determines marketplace welfare when agents exhibit first-proposal bias.