When AI summaries appear in search results, the likelihood of users clicking on regular web pages is nearly halved. An analysis by the Pew Research Center of large-scale browsing behavior among U.S. users on Google found that the proportion of users clicking on traditional search results dropped directly from 15% to 8% when AI summaries were present. [1]
Users are completing exploration and comparison within conversations, and ads have migrated along with them. OpenAI has officially launched ChatGPT Ads, and Google has introduced ads into AI Overviews, attaching sponsored cards to responses—directly continuing the model from the search engine era. [2, 3] But as the paradigm of how people acquire information shifts, it prompts a rethinking of advertising formats:
When the medium of information shifts from a full page of search results to a real-time generated response, what should advertising auction?
For thirty years, online advertising has been built on a default assumption: ad slots exist first, and auctions determine who wins them. The ranking of search results or which ad appears in a feed are pre-defined opportunities. But large model responses are generated word by word: whether a brand should appear, where it should appear, and how it should be presented all depend on what has already been written. Even whether to discuss price or experience first can create entirely different impressions of the product.

Figure 1: In generative information media, where advertising opportunities arise, who captures them, and how they are presented collectively influence marketing effectiveness and user experience.
Thus, generation does more than fill ad opportunities—it shapes them. Under this new paradigm, existing research has begun exploring auction mechanisms such as paragraph-level ad insertion and candidate response bidding, while a research team from Renmin University’s Gaoling School of Business and Stanford University has further advanced native large model advertising by extending competition into the token-by-token generation process.
Their proposed LAMA enables content generation and ad allocation to evolve together: every time a token is generated, it simultaneously influences how the response continues and determines who ultimately receives ad exposure. This is the core idea introduced by this work—Generation as Allocation.

Token-Level Advertising
Authors: Hanbing Liu, Bowei Zhang, Changyuan Yu, Yinyu Ye, Qi Qi
Affiliation: Huiyan School of Artificial Intelligence, Renmin University of China; Stanford University
Link: https://arxiv.org/abs/2608.27382
The large model writes one character at a time.
It is all a single auction process.
Advance ad auctions to the token level, so advertisers are no longer just competing for final exposure opportunities, but also for control over what comes next. For example, when recommending a vacation destination, deciding whether to mention transportation first or accommodation first could give different advertisers an advantage in subsequent competition. This gives advertisers new strategic flexibility: they can not only adjust their bids to influence the current output, but also change their future strategies in real time based on what has already been generated.
This places higher demands on the mechanism: not only must advertisers be willing to provide truthful value information at the outset, but they must also continue to tell the truth throughout the generation process. At the same time, responses must not solely cater to bids and deviate from the user’s original question.
LAMA stands for Latent Advertiser Mixture Auction. It enables advertisers to report, for the current text, “How valuable would it be to me if the next part were written this way?” These reports are combined with the original language model to form generation strategies reflecting each advertiser’s preferences. The platform then generates the response token by token, continuously updating the probability that each advertiser will ultimately receive exposure: the more a newly generated segment aligns with an advertiser’s preferred direction, the higher its allocation probability becomes. After the response is complete, the winner is selected based on the final probabilities and granted exposure in the form of external links, sponsored cards, or similar formats. In practical deployment, these value reports can be computed by platform-side models, analogous to today’s auto-bidding services in advertising systems.

Figure 3: The LAMA process. During each token generation, the steps of reporting, generating, and updating are completed sequentially, followed by final allocation and settlement once the answer is fully generated.
How can we prevent advertisers from reneging on this process? LAMA has carefully designed accompanying local report verification and payment rules for this generation process.
Verification focuses on ensuring consistency between prior and subsequent reports: the advertiser’s value assessment of different next-step variations must align with the earlier report on the current text. The platform only needs to perform a local check at each step to prevent advertisers from arbitrarily changing their positions. Payments are adjusted in tandem with the generation process: if a step increases the advertiser’s probability of exposure combined with subsequent value, a corresponding fee is charged; if it reduces the overall benefit, a subsidy is provided, and all amounts are settled together with the initial fee. This links influence over generation, pursuit of exposure, and cost responsibility together.
This mechanism satisfies Markov Dominant Strategy Incentive Compatibility (Markov DSIC): regardless of what feasible strategies other advertisers adopt, if an advertiser has been truthful in prior reports, then continuing to report truthfully at any reachable generated state is no worse than switching to any other feasible strategy thereafter. In other words, honesty applies not only to initial bids but also throughout the entire generation process.
Meanwhile, LAMA also includes social welfare guarantees: by targeting the advertising value minus a penalty for deviation from native responses, it has a well-defined theoretical upper bound on its gap from the ideal optimal mechanism. This ensures that while pursuing commercial value, the mechanism also limits the impact of advertising on user experience.
From how each step is written to who ultimately receives exposure and how much is paid, LAMA incorporates all of these into a single mechanism. This is precisely what “generate-to-allocate” means.
The theory is sound.
How effective is it in practice?
The research team selected 1,239 real commercial search queries from the Webis Generated Native Ads 2024 dataset for a proof of concept, covering three common scenarios: fitness, vacation, and automobiles, each representing a three-advertiser competitive environment.
The experiment uses Qwen3-14B as the reference language model, with an advertiser-shared report model trained via LoRA. Advertiser value is derived from estimates predicted by an open-source predictive model, while user experience is comprehensively evaluated based on GEM-Bench metrics including relevance, informativeness, fluency and coherence, ad naturalness, ad factuality, and intrusiveness. Baseline methods include two assignment strategies before and after generation, combined with three generation approaches—native content, edited insertion, and reinforcement learning—and the answer-level aggregation mechanism MOSAIC.

Figure 4: Main Experiment Table
LAMA achieved the highest scores across all four metrics, successfully delivering mutual gains in platform revenue (+10.7%) and advertiser value (+3.8%) without compromising user experience at all.
Let’s examine a specific query in the vacation context: “best beach vacations in the world.” A reference answer without advertising mechanisms discusses only destinations and planning factors; the LAMA-generated response mentions Expedia in the first part and Tripadvisor in the latter, with advertisers alternately taking control during generation—yet the output remains natural and fluent, providing valuable information just like the reference answer.

Figure 5: Sample response in a vacation scenario; the orange-highlighted text segment is assigned to the hyperlink.
After generation, Tripadvisor had a significantly higher allocation probability and was selected as the final winner in this response.
Why Tripadvisor? From the text, Expedia seems to serve more as a brand anchor in the first half to introduce destinations like the Maldives, Bali, and the Caribbean; the sentence mentioning Tripadvisor clearly articulates what the product does: compare destinations, read traveler reviews, and filter accommodations and experiences—offering a more complete value proposition. The platform’s allocation probability update keenly captured this.
The research team also conducted an intuitive bid perturbation experiment: keeping all other advertisers' reports fixed, they uniformly adjusted the target advertiser's report up or down from its true value. The results aligned closely with intuition: the higher the bid, the greater the token share, allocation probability, and payment. However, the true expected utility does not increase indefinitely—it peaks precisely at truthful reporting. While inflating bids can indeed secure more influence in the outcome, the increased payment outweighs the benefit, exactly as guaranteed by the theoretical property of "incentive compatibility."

Figure 6: Generation share, allocation, and payments all increase with reporting, and the true expected utility peaks when reporting is truthful.
Conclusion
Where are generative ads most suitable? A natural fit is in scenarios where users are already seeking commercial advice: comparing products using digital shopping assistants, selecting hotels while traveling, or letting an agent choose paid tools for daily workflows. When commercial information responds to users’ actual needs, clearly labeled as advertising, and does not compromise the validity or objectivity of the response, it is more likely to be accepted.
LAMA explores how to implement the mechanisms under this vision: enabling advertisers to compete during the generation process while using rules to constrain strategic behavior and balance commercial value with answer quality.
In the past, auctions determined who obtained an existing ad slot. Today, the mechanism can also determine how ad opportunities are formed within responses. Moving from “allocating positions” to “generating while allocating,” LAMA offers a new design path for native advertising mechanisms in large models.
Author Information
Liu Hanbing (first author): A third-year master’s student at the Huiyuan School of Artificial Intelligence, Renmin University of China, supervised by Professor Qi Qi. Her research focuses on intelligent decision-making and mechanism design in generative AI, with an expected graduation in 2027. Her work has been published at top-tier conferences including WINE, ACL, and SIGMOD.
Homepage: https://tokgoleo.github.io/home/
Qi Qi (Corresponding Author): Tenured Professor at the Huiyuan School of Artificial Intelligence, Renmin University of China; National High-Level Young Talent; Deputy Director (Executive) and Director of the Office of the Special Committee on Computational Economics, China Computer Federation (CCF). Ph.D. from Stanford University, advised by Professor Yinyu Ye. Research interests include algorithmic game theory and mechanism design, large models and agents, multi-agent systems, optimization, and intelligent decision-making. His work has been widely published in top-tier journals such as MS, OR, MOR, GEB, and TR-B, as well as CCF-A class conferences including STOC, ICML, NeurIPS, ICLR, ACL, AAAI, KDD, SIGIR, and WWW. He has also maintained long-term collaborations with leading enterprises such as Baidu, Alibaba, Tencent, Meituan, ByteDance, Didi, Huawei, and OPPO, driving the real-world deployment of large models, mechanism design, and multi-agent decision technologies in production environments serving hundreds of millions of users, with applications in online advertising, intelligent marketing, coupon optimization, and user growth.
Homepage: https://ai.ruc.edu.cn/szdw/zrjs/qq/index.htm
Reference materials
1. Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results, 2025-07-22: https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
2. OpenAI Help Center, Ads in ChatGPT, accessed 2026-09-04: https://help.openai.com/en/articles/20001047-ads-in-chatgpt
3. Google Ads & Commerce, More opportunities for your business on Google Search, 2025-05-21: https://blog.google/products/ads-commerce/google-search-ai-brand-discovery
This article is from the WeChat public account "Machine Heart".
