Problem

The emergence of conversational AI, particularly Large Language Models (LLMs), is transforming how users interact with online platforms to find information. Instead of traditional search results, users now receive direct answers and clarifying follow-up prompts. This shift presents a unique opportunity to address a major challenge in advertising: ambiguous user intent. For instance, a user searching for "running shoes" might have vastly different needs (e.g., trail vs. road, beginner vs. experienced). While some existing search features offer refinements, LLM-generated sponsored questions can proactively clarify user intent. The central problem is how to efficiently and fairly allocate these "suggestion slots" to advertisers and how this new mechanism integrates with subsequent ad auctions. A critical design choice arises: should platforms build an end-to-end system that jointly optimizes both the sponsored suggestion and the ad auction, or should they implement simpler, decoupled modular mechanisms for each stage?

Method

The authors develop a formal model for advertising in interactive platforms, combining information elicitation and ad allocation. This model conceptualizes user intent as an unknown state of nature, advertisers as having private base values, and publicly known state-dependent conversion rates. A "question" posed by the platform generates a signal, which updates all parties' beliefs about the user's latent intent. The study analyzes two primary design philosophies:

1. End-to-End Mechanism: This approach involves a single, direct revelation mechanism where advertisers report their private base values. The platform then uses this information to jointly decide which sponsored question to present, which ad to allocate after a signal is observed, and the payments. The paper demonstrates that a welfare-maximizing mechanism can be instantiated as an end-to-end Vickrey-Clarke-Groves (VCG) auction. This VCG mechanism is proven to be dominant strategy incentive compatible (DSIC), meaning advertisers are incentivized to report their true values. The VCG payment can be decomposed into an externality from the question selection stage and an expected second-price payment from the ad allocation stage.

2. Modular Two-Stage Mechanism: This alternative decouples the problem into two separate stages. First, an auction is run to select a sponsored question. Second, after the signal from the chosen question is observed, another auction is run to allocate the final ad slot. The authors specifically consider a "VCG-per-stage" design, where both stages use VCG-like auctions. In this setup, advertisers bid their expected utilities for each question in the first stage and their true base values for the item in the second stage, anticipating the truthful and efficient outcome of the second stage. This is shown to constitute a pure Nash equilibrium.

Results

The analysis yields crucial findings regarding the effectiveness of each design approach:

1. The end-to-end VCG mechanism is proven to be truthful and maximizes expected social welfare, ensuring efficient and desirable outcomes for the platform and advertisers.

2. In contrast, the seemingly simpler modular VCG-per-stage mechanism, despite having a natural pure Nash equilibrium, suffers from severe strategic inefficiency. The paper proves that the Price of Anarchy for this modular design is unbounded. This implies that the social welfare achieved by the modular approach can be arbitrarily worse than the optimal outcome attainable by a unified, end-to-end mechanism.

3. The unbounded Price of Anarchy is robust and persists even if the first-stage auction of the modular mechanism uses other common payment formats, such as first-price or all-pay auctions, while retaining the "highest sum of bids" allocation rule for questions.

Implications

The findings strongly advocate for a unified, end-to-end mechanism for allocating sponsored suggestions in interactive platforms, particularly within conversational AI contexts. While modular designs may appear simpler to implement, they introduce significant strategic inefficiencies, leading to potentially sub-optimal outcomes. The paper highlights that a unified VCG mechanism offers both truthfulness and efficiency guarantees, making it a more attractive and robust solution for complex advertising ecosystems where user intent is interactively clarified. Future research could explore mechanisms designed for revenue maximization rather than welfare maximization, and investigate hybrid systems that integrate both organic and sponsored interactions over multiple conversational rounds, considering the complex dynamics and policy implications involved.