Ad Insertion in LLM-Generated Responses
Shengwei Xu, Zhaohua Chen, Xiaotie Deng, Zhiyi Huang, Grant Schoenebeck
TL;DR
This paper tackles sustainable monetization of LLMs by embedding ads into responses in a way that preserves contextual coherence and minimizes latency while respecting privacy and disclosure rules. It introduces a two-layer decoupled framework: (i) insertion is decoupled from response generation, ensuring explicit ads; (ii) bidding is decoupled from specific queries via a fixed set of genres, enabling offline bids and online slot-coherence estimation. The core mechanism is a VCG auction over $L$ genres that assigns $K$ ads per response using proxy genre-values $\tilde{v}_{i,g}$ and a coherence vector $\tilde{c}_{j}$, with theoretical guarantees of approximately DSIC and IR (error bounded by $\varepsilon = \varepsilon_V + \varepsilon_C$) and near-optimal social welfare; the welfare gap is bounded by $2K \cdot \varepsilon \cdot \bar{v}$. To estimate coherence, the authors propose two methods (sentence embeddings and LLM-as-a-Judge) and validate them against human judgments, finding that LLM-as-a-Judge (e.g., GPT-5) achieves a Spearman correlation of about $\rho \approx 0.66$ with human ratings, outperforming most humans. Prototypes demonstrate real-time viability: VCG runs within seconds even for large advertiser pools and abundant slots, supporting practical deployment. Altogether, the work offers a scalable, privacy-aware path to monetizing LLM services with coherent, explicitly disclosed ads and robust welfare guarantees.
Abstract
Sustainable monetization of Large Language Models (LLMs) remains a critical open challenge. Traditional search advertising, which relies on static keywords, fails to capture the fleeting, context-dependent user intents--the specific information, goods, or services a user seeks--embedded in conversational flows. Beyond the standard goal of social welfare maximization, effective LLM advertising imposes additional requirements on contextual coherence (ensuring ads align semantically with transient user intents) and computational efficiency (avoiding user interaction latency), as well as adherence to ethical and regulatory standards, including preserving privacy and ensuring explicit ad disclosure. Although various recent solutions have explored bidding on token-level and query-level, both categories of approaches generally fail to holistically satisfy this multifaceted set of constraints. We propose a practical framework that resolves these tensions through two decoupling strategies. First, we decouple ad insertion from response generation to ensure safety and explicit disclosure. Second, we decouple bidding from specific user queries by using ``genres'' (high-level semantic clusters) as a proxy. This allows advertisers to bid on stable categories rather than sensitive real-time response, reducing computational burden and privacy risks. We demonstrate that applying the VCG auction mechanism to this genre-based framework yields approximately dominant strategy incentive compatibility (DSIC) and individual rationality (IR), as well as approximately optimal social welfare, while maintaining high computational efficiency. Finally, we introduce an "LLM-as-a-Judge" metric to estimate contextual coherence. Our experiments show that this metric correlates strongly with human ratings (Spearman's $ρ\approx 0.66$), outperforming 80% of individual human evaluators.