Jointly auctioning sponsored follow-up prompts with downstream ads using a VCG-style mechanism yields efficient, incentive-compatible outcomes; by contrast, a simple modular design that separates suggestions from the ad auction can be strategically exploited and lead to unbounded inefficiency.

Main Finding

A unified, end-to-end VCG-style mechanism that jointly chooses which LLM-generated clarifying suggestion to show and runs the downstream ad allocation is both efficient and truthful (DSIC). In contrast, a natural modular design that auctions the suggestion slot separately from the subsequent ad auction (e.g., VCG-per-stage, first-price, or all-pay for the suggestion) admits equilibria with unbounded Price of Anarchy—i.e., arbitrarily bad welfare relative to the end-to-end optimum—and faces serious practical strategic and informational difficulties.

Key Points

• Problem setup

  • Platforms can present one of k possible sponsored follow-up prompts (questions) before allocating a single ad slot.
  • A latent user state θ (sampled from public prior D) determines ad relevance via public conversion rates αi(θ); each advertiser i has a private base value vi and value for allocation = vi · αi(θ).
  • Asking question ℓ produces a publicly observed signal σ drawn from Qℓ(· | θ); after observing σ all players update to posterior Dℓ,σ and the platform runs the ad allocation.
  • Platform objective: maximize expected welfare (sum of advertiser values given posterior).

• End-to-end (direct revelation) approach

  • The platform asks advertisers to report base values once, then chooses a suggestion, observes the signal, allocates the ad, and charges payments.
  • The authors show a VCG-style mechanism can be implemented end-to-end: it is dominant-strategy incentive compatible (truthful) and welfare-maximizing.
  • The VCG payment decomposes into two parts: (1) a familiar second-price component from the final ad allocation, and (2) an additional payment capturing the marginal impact of an advertiser’s reported bid on the choice of question (the information-acquisition decision).

• Modular (two-stage) approach and equilibrium characterization

  • Natural modular design: stage 1 auction chooses the question (advertisers bid their expected continuation value), stage 2 auction sells the ad slot after signal realization.
  • There exists an intuitive pure Nash equilibrium of the VCG-per-stage mechanism where advertisers bid in stage 1 their expected utility from the subsequent ad auction and bid truthfully in stage 2.
  • However, this equilibrium is strategically fragile and practically problematic: computing a stage-1 value requires knowledge of others’ values/strategies; relying on the platform to compute these (“proxying”) gives incentives to misreport; bids on standalone questions may miss context dependence of conversational flows.

• Efficiency failure of modular designs

  • The equilibrium welfare of the VCG-per-stage modular mechanism can be arbitrarily worse than the end-to-end optimum (unbounded Price of Anarchy).
  • The inefficiency is not limited to VCG-per-stage: analogous pathological outcomes occur when first-price or all-pay auctions are used in the suggestion stage.

• Illustrative example

  • The "running shoes" example shows how different questions (e.g., “trail running?” vs “experienced trail runner?”) induce different posterior distributions and hence different expected advertiser values; it demonstrates how the choice of suggestion meaningfully affects allocation efficiency and payments.

Data & Methods

• Nature of work: theoretical mechanism-design analysis (no empirical datasets).
• Formal model:
  • Single ad slot, finite state space Θ, public prior D over θ.
  • k candidate questions; each question ℓ has signal space Σℓ and conditional distributions Qℓ(σ | θ).
  • Advertisers: private base values vi ≥ 0; public conversion rates αi(θ) ≥ 0; quasilinear utilities.
  • Timing: θ drawn → (advertisers report in direct mechanism or bid in stage 1 for modular) → platform picks question → signal realized and observed → stage-2 bidding or allocation → payment.
• Methods:
  • Construct an end-to-end direct-revelation VCG mechanism adapted to this information-acquisition + allocation environment; prove DSIC and characterize payments (including decomposition).
  • Analyze modular two-stage mechanisms: identify natural equilibria (VCG-per-stage), examine computability and strategic issues (proxy manipulation), and prove formal lower bounds on welfare (unbounded Price of Anarchy).
  • Use constructive examples and worst-case proof techniques to show inefficiency persists across common auction formats for the suggestion stage.

Implications for AI Economics

• Mechanism design for conversational/LLM-driven ad placements must account for information acquisition effects. Suggestions (clarifying prompts) change posteriors and thus downstream ad valuations—so treating them independently risks grave inefficiency.
• End-to-end auctions that jointly optimize information elicitation (which suggestion to show) and allocation are theoretically preferable: they can restore truthfulness and welfare-optimality. Platforms building conversational ad products should consider integrated mechanisms, particularly when automated bidding agents (autobidders) are used and can report/value-compute on advertisers’ behalf.
• Practical trade-offs:
  • End-to-end VCG is conceptually and computationally more complex (requires the platform to compute counterfactuals about question choice and downstream outcomes), but it avoids strategic pathologies inherent in modularization.
  • Modular designs are simpler to implement but expose platforms and advertisers to large welfare losses and strategic manipulation; they also impose unrealistic informational burdens on advertisers (to compute context-sensitive stage-1 values).
• Policy and product design considerations:
  • Platforms should be cautious about allowing advertiser influence over conversational steering (sponsored prompts) because it raises user-experience and regulatory concerns; the paper frames its results as technical guidance assuming the platform elects to pursue sponsored suggestions.
  • If sponsored suggestions are deployed, platforms should favor mechanisms that internalize the information-acquisition externality (joint design) and support advertiser-side automation (autobidders) to manage complexity.
• Directions for future AI-economics work:
  • Extend analysis to multiple ad slots, dynamic/multi-turn dialogues, private conversion rates, or richer advertiser information structures.
  • Empirical and simulation studies to evaluate trade-offs between implementational complexity and welfare gains in realistic conversational-ad environments.
  • Study robustness to misspecification of priors, learning dynamics, and the impact of policy constraints prioritizing user experience over revenue.