Generation as Allocation: LAMA Embeds Ad Auctions Into LLM Token Generation

AI-generated summaries are dramatically reducing clicks on traditional search results. Pew Research Center found that when an AI summary appears, user clicks on organic links drop from 15% to 8%. Meanwhile, OpenAI and Google have begun placing sponsored cards inside ChatGPT and AI Overviews, extending the classic search-ad model into generative answers.

For three decades online advertising has assumed a fixed inventory: ad slots exist first, then auctions decide who fills them. LLM answers, however, are generated token by token; whether a brand appears, where it appears, and how it is framed all depend on the preceding context. The order of arguments — price first vs. experience first — can completely reshape product perception.

This insight motivates LAMA (Latent Advertiser Mixture Auction), a mechanism that unifies content generation and ad allocation. At each generation step, advertisers report the value of continuing the answer in different directions. The platform combines these reports with the base language model to produce the next token, then updates each advertiser's probability of winning the final exposure. After the answer finishes, the winner is selected according to the final probabilities and receives a sponsored link or card. In practice, value reports can be computed by a platform-side model, analogous to today's auto-bidding systems.

To prevent advertisers from changing their reports mid-generation, LAMA introduces local report verification and a payment rule tied to the generation path. Verification checks that an advertiser's stepwise value reports remain consistent with its earlier report for the current prefix. Payments adjust at each step: if a token increases an advertiser's allocation probability and future value, a charge accrues; if it decreases them, a subsidy is granted. The sum of initial and stepwise payments settles at the end.

The mechanism satisfies Markov Dominant-Strategy Incentive Compatibility (Markov DSIC) : at any reachable generation state, truthful reporting is a dominant strategy regardless of other advertisers' strategies. Honesty is therefore sustained throughout the entire generation, not only at the outset.

LAMA also provides a social welfare guarantee . The objective maximizes advertiser value minus a penalty for deviating from the native answer, and the gap to the ideal optimum has a proven theoretical upper bound. This constrains the commercial distortion of user experience.

Experimental Validation

The team evaluated LAMA on 1,239 real commercial queries from the Webis Generated Native Ads 2024 dataset, covering fitness, vacation, and automotive scenarios — each with three competing advertisers. They used Qwen3-14B as the base language model, trained a shared reporting model via LoRA, estimated advertiser value with an open-source predictor, and assessed user experience using GEM-Bench metrics (relevance, informativeness, fluency, ad naturalness, factuality, intrusiveness). Baselines included pre-generation and post-generation allocation combined with native generation, editorial insertion, RL-based generation, and the answer-level aggregation mechanism MOSAIC.

LAMA achieved the highest scores on all four metrics: platform revenue (+10.7%), advertiser value (+3.8%), user experience (no degradation), and allocation efficiency.

Case Study: Vacation Query

For the query "best beach vacations in the world", the native answer discussed destinations and planning factors. LAMA's generated answer mentioned Expedia in the first half and Tripadvisor in the second half, with control alternating between advertisers. The final allocation probability favored Tripadvisor because its associated text delivered a more complete value proposition — comparing destinations, showing traveler reviews, filtering accommodations and experiences — while Expedia served mainly as a brand anchor for destination names. The generated answer remained natural, fluent, and informative.

Bid Perturbation Experiment

Holding other advertisers' reports fixed, the researchers scaled a target advertiser's reports up and down from truthful values. Token contribution share, allocation probability, and payment all increased with higher reports, but the advertiser's true expected utility peaked exactly at honest reporting. Overbidding buys more influence but incurs disproportionately higher payments, confirming the incentive-compatibility property.

Conclusion

Generative advertising fits naturally where users seek commercial advice: product comparison, travel planning, tool selection. When commercial information answers a genuine need, carries clear ad labels, and preserves factual integrity, acceptance rises. LAMA demonstrates how to realize this vision: advertisers compete during generation while rules constrain strategic behavior and balance commercial value with answer quality. The paradigm shifts from "auctioning pre-existing slots" to "mechanisms that shape how ad opportunities form inside the answer" — from allocating positions to generation as allocation .

Paper: https://arxiv.org/abs/2608.27382 (Hanbing Liu, Bowei Zhang, Changyuan Yu, Yinyu Ye, Qi Qi; RUC Gaoling School of AI, Stanford University).