Agentic Markets: Equilibrium Effects of Improving Consumer Search

1. Problem Statement

The paper investigates the long-term economic and learning dynamics of agentic markets—two-sided markets where AI agents assist consumers in searching for products and businesses. While AI agents can reduce search costs and increase the informativeness of pre-purchase inquiries (e.g., by allowing adaptive questioning rather than rigid filters), the authors ask: How do these improvements affect market-level learning, competition, and consumer welfare in the long run?

The core tension lies in the difference between individual transaction benefits (where better search helps a specific consumer) and aggregate market outcomes (where information generated by early searches shapes the beliefs and choices of future consumers). The authors specifically analyze two dimensions of improvement:

1. Cheaper Search: Reducing the cost ($c$) of inspecting a product.
2. More Informative Search: Increasing the depth of information revealed before a transaction (e.g., identifying specific reasons for a "fit" mismatch).

2. Methodology and Model

The authors propose a tractable, sequential model of consumer search and market learning.

• Market Structure:

  • An infinite stream of consumers arrives sequentially.
  • A fixed pool of $m$ businesses offers products at prices $P_j$.
  • Utility: A consumer $i$ derives value $V_i$ if a transaction meets their needs. The need is met only if two conditions are satisfied:
    1. Fit ($F_{ij}$): A binary attribute (e.g., cuisine type, dietary restrictions) that can be inspected before purchase at cost $c$.
    2. Quality ($q_{ij}$): A binary attribute (e.g., taste, reliability) revealed only after the transaction.
  • Parameters: Each business $j$ has an unknown true fit probability $\pi_j$ and quality probability $Q_j$. Consumers hold Bayesian priors over these parameters.

• Consumer Search Policy:

  • Modeled as a variant of Weitzman's Pandora's Box problem.
  • Consumers use an optimal search policy (Algorithm 1) based on an index $\sigma_{ij} = V_i \hat{Q}_{ij} - P_j - c/\hat{\pi}_{ij}$, where $\hat{\pi}$ and $\hat{Q}$ are posterior means.
  • Consumers inspect businesses in decreasing order of this index. If a business passes the fit check ($F_{ij}=1$), the consumer transacts immediately.

• Market Dynamics & Learning:

  • After each round, the platform collects feedback: which businesses were inspected, their fit outcomes, and the quality outcomes of transactions.
  • This history updates the posterior beliefs of future consumers.
  • Steady State: The market converges to a state where businesses are either Learned (parameters $\pi_j, Q_j$ are known to the market) or Lost (posterior beliefs drop so low that no consumer ever inspects them again).

• Extensions:

  • Sequential Screening (Informativeness): The model is extended to "screen sequences" where a search reveals a sequence of binary signals. Increasing the "search degree" $k$ makes search more informative (revealing more specific reasons for failure) but potentially more selective.
  • Transcripts: A mechanism where the platform observes why a fit failed (e.g., "vegan menu missing" vs. "no insurance"), allowing the market to distinguish between niche mismatches and broad quality deficiencies.
  • Endogenous Pricing: Businesses are modeled as strategic agents who adjust prices to maximize revenue in response to consumer search behavior.

3. Key Contributions

1. Novel Model of Agentic Markets: Introduces a framework combining sequential search (Weitzman), social learning (bandit dynamics), and strategic pricing to analyze AI-mediated markets.
2. Decoupling Cost vs. Informativeness: Distinguishes between the effects of lowering search costs and increasing search depth, showing they have divergent impacts on long-run welfare.
3. The Role of Transcripts: Identifies that the negative effects of highly informative search can be mitigated if the platform aggregates "transcripts" (reasons for failure) rather than just binary pass/fail outcomes.
4. Equilibrium Analysis with Endogenous Prices: Characterizes market equilibria in symmetric markets, showing how search improvements alter competitive pressure and pricing strategies.

4. Key Results

A. Exogenous Prices (Fixed Prices)

• Cheaper Search: Unambiguously improves long-run learning and consumer surplus. Lower costs encourage exploration, preventing viable businesses from being "lost" due to insufficient data.
• More Informative Search (without Transcripts): Can degrade long-run learning and consumer surplus.
  • Mechanism: Highly informative search reveals specific, idiosyncratic mismatches (e.g., "wrong color"). If the market cannot distinguish these from broad quality failures, it treats all failures as evidence of low quality. This causes broadly appealing businesses to be prematurely abandoned ("lost") because they fail specific, narrow checks.
• More Informative Search (with Transcripts): If the platform observes the reason for failure (transcripts), increased informativeness becomes unambiguously beneficial. The market learns that a failure was due to a niche mismatch rather than poor quality, preserving the business for future consumers.

B. Endogenous Prices (Strategic Pricing)

• Cheaper Search: Continues to benefit consumers. Lower costs increase the set of learned businesses, intensifying competition among firms, which drives prices down and increases total welfare.
• More Informative Search (with Transcripts): Can reduce consumer surplus even if learning improves.
  • Mechanism: Highly informative search acts as a filter, narrowing the set of businesses that "pass" for a specific consumer. This reduces the effective competition faced by the remaining businesses. With fewer rivals, firms can charge higher prices, extracting more surplus from consumers.
  • Result: There exist symmetric market equilibria where increased search informativeness leads to strictly lower consumer welfare due to weakened competitive pressure.

C. Symmetric Market Equilibrium

• The authors characterize the unique symmetric equilibrium for learned-symmetric markets.
• They prove that while cheaper search increases total welfare (consumer utility + firm revenue), the distribution of gains depends on the number of surviving firms.
• They demonstrate that in some cases, firms might prefer higher search costs because high costs eliminate competitors, allowing the survivors to charge monopoly-like prices.

5. Significance and Implications

• Platform Design: The paper highlights that simply making search "smarter" or "deeper" is not sufficient for welfare maximization. Platforms must design information aggregation mechanisms. Specifically, capturing and sharing transcripts (reasons for rejection) is crucial to prevent the market from misinterpreting niche mismatches as quality failures.
• Competition vs. Matching: There is a trade-off between matching efficiency (finding the perfect product) and competitive intensity. Highly precise matching can reduce the number of viable options for a consumer, thereby reducing competition and allowing firms to raise prices.
• Policy & AI Deployment: As AI agents become standard, regulators and platform designers must consider that improvements in individual user experience (e.g., more detailed AI questioning) could inadvertently harm aggregate market efficiency and consumer surplus if they alter the competitive landscape or information aggregation negatively.
• Theoretical Foundation: Provides a rigorous mathematical basis for future research on agentic markets, moving beyond static models to dynamic, learning-based equilibria.

Summary Conclusion

The paper concludes that cheaper search is universally beneficial, but more informative search is a double-edged sword. Without proper information aggregation (transcripts), it harms learning; with transcripts, it improves learning but may harm consumer surplus by reducing competition. The ultimate welfare outcome depends critically on how platforms observe, aggregate, and share the rich data generated by agentic interactions.