Smarter search can hurt consumers: more informative AI-assisted search may reduce learning and consumer surplus and soften price competition unless markets themselves learn from agent-consumer interactions; by contrast, cheaper search unambiguously benefits consumers.

Main Finding

Improving consumer search via AI agents has two distinct and opposing equilibrium effects. Lowering search costs unambiguously increases exploration, improves market-level learning, and raises consumer surplus. By contrast, making individual searches more informative (shifting information from post-purchase to pre-purchase) can paradoxically reduce long-run learning and consumer welfare unless platforms also surface the richer, diagnostic information (e.g., “transcripts” of why a business failed a check). When firms set prices endogenously, cheaper search strengthens competition and helps consumers, while more informative search can weaken competition (by narrowing each consumer’s viable set) and thus raise prices, reducing consumer surplus even if learning improves.

Key Points

• Model intuition

  • Each transaction value depends on two binary components: fit (inspectable at cost) and quality (revealed only after purchase). Consumers pay cost c to inspect fit; quality is learned only upon transaction.
  • Consumers arrive sequentially, observe the full history of past inspections and transactions, and choose inspections to maximize expected utility. Businesses are fixed in number, with (unknown) fit and quality parameters (πj, Qj).
  • The consumer search problem reduces to a Weitzman-style index policy: inspect businesses in decreasing order of σij := Vi·Q̂ij − Pj − c/π̂ij and transact with the first inspected business that fits.

• Main theoretical results

  • Exogenous prices:
    • Reducing search cost c increases exploration: more businesses become “learned” (posterior beliefs converge to true parameters), improving long-run consumer surplus.
    • Increasing informativeness of search (more pre-purchase revelation of fit details) can reduce market-level learning and consumer surplus. Mechanism: failures during inspection become dominated by idiosyncratic, consumer-specific mismatches; if the market only observes a binary pass/fail (without reason), it cannot distinguish niche mismatches from broad defects and may prematurely abandon otherwise broadly appealing businesses.
    • If the platform records and shares diagnostic transcripts (which failure item caused the rejection), increased informativeness becomes unambiguously beneficial for learning and welfare.
  • Endogenous prices:
    • Equilibrium characterization in symmetric markets (building on Choi et al., 2018) shows cheaper search intensifies competition and benefits consumers.
    • More informative screening can reduce consumer surplus via a competition channel: tighter pre-purchase screening shrinks the set of firms each consumer considers viable, weakening competitive pressure and allowing higher equilibrium prices even when learning is better.

• Mechanism illustration

  • Catering example: agentic search can let consumers probe many fit dimensions (dietary constraints, availability). If failures are aggregated as undifferentiated “didn’t pass”, a broadly good caterer might be abandoned because many consumer-specific rejections mask its general quality. If platforms record which requirement caused the failure, that information keeps the market from abandoning such firms.

• Important assumptions and caveats

  • Independence of fit and quality across businesses is central to analysis; some relaxations are possible but would change dynamics.
  • Binary signals (fit yes/no, quality yes/no) simplify exposition; richer signal spaces are an avenue for extension.
  • Platform perfectly records and shares whatever feedback the model assumes (the distinction between pass/fail vs transcript matters crucially).

Data & Methods

• Methods: Analytic/theoretical model and equilibrium analysis.
  • Sequential Bayesian model of consumer search: an infinite sequence of consumers, fixed set of m businesses, consumers drawn i.i.d. values Vi from distribution DV.
  • Each consumer can pay cost c to inspect fit Fij (binary), and only a transaction reveals binary quality qij; priors over business parameters (πj, Qj) are common and updated via Bayes using the full public history.
  • Consumers follow optimal inspection order derived from Weitzman’s Pandora’s box solution (index policy σij).
  • Long-run behavior analyzed via steady-state limits: which businesses are “learned” (posteriors converge to true parameters) and how expected consumer utility behaves as c and informativeness vary.
  • Two policy regimes studied: exogenous fixed prices, and endogenous price-setting by firms (symmetric equilibrium analysis drawing on Choi et al., 2018 techniques).
• No empirical data; results are theoretical with proofs (convergence, comparison statics) and constructed counterexamples demonstrating potential welfare reversals when informativeness increases without transcript-sharing.

Implications for AI Economics

• Platform design matters fundamentally. Agentic capabilities change not only per-consumer match quality but also the information aggregation process that shapes future consumers’ behavior.
  • Make search cheaper: Policies, product design, or agentic features that lower the per-search cost (e.g., automated multi-site probing) are robustly pro-consumer: they increase exploration, market learning, and competition.
  • Make search informative with care: Increasing pre-purchase information capture can hurt aggregate learning and consumer surplus if platforms only record coarse outcomes (pass/fail). Platforms should design mechanisms to capture and share diagnostic details (transcripts, failure reasons, structured attributes) so richer searches translate into better market-level learning.
• Competition and pricing tradeoffs: Even if consumer-level matches improve with more informative search, the narrowing of viable firm sets per consumer may soften competition and raise prices. Regulators and platform designers should consider this second-order effect; interventions could include:
  • Encouraging broader exposure or subsidized exploration to counteract firms’ ability to price higher when screening narrows consideration sets.
  • Standardizing the representation of diagnostic signals so consumers and the market can distinguish niche mismatches from broad defects.
• Future research priorities suggested by the paper:
  • Relaxing independence assumptions and allowing correlated fit/quality across firms or consumers.
  • Modeling richer (non-binary) signals, noisy transcripts, or partial transcript access.
  • Considering strategic behavior by firms beyond price (e.g., signaling, manipulating transcript content, or agent-to-agent protocols).
  • Empirical validation in domains where agentic search is being deployed (marketplaces, travel, hiring platforms) to measure how agentic transcripts and search costs affect exploration and price dynamics.

Summary takeaway: Agentic search tools are not unambiguously welfare-improving at the market level — lowering search costs is reliably good, but increasing pre-purchase informativeness requires careful platform practices (notably sharing diagnostic information) to avoid unintended reductions in learning and consumer surplus, especially once firms set prices strategically.