Google explores auction model where brands bid to shape AI-generated ad responses
Instead of bidding for ad space, advertisers could bid to shape the AI’s words, token by token
In traditional Google Ads, advertisers bid on keywords, submit a fixed piece of creative, and hope for prime placement. That model has shaped digital advertising for years. But according to new research from Google Research and the University of Chicago, the future could look very different and far more dynamic.
Search Engine Land explains the new system. The research introduces a system where advertisers don’t just bid to place an ad; they bid to shape how a large language model (LLM) responds to a user’s query. Instead of selecting from a pool of static creatives, the system would generate the ad in real time, token by token, influenced by the brands involved.
How the Token Auction Model works
In this proposed system, called the Token Auction model, every advertiser submits a bid along with their own custom-tuned LLM. This model captures the brand's voice, including its preferred tone, phrases, and messaging style. When a user enters a query, the generative system writes a response one word at a time, weighing how much each advertiser's model should influence the next token based on the strength of their bid.
Let’s break it down:
- Each token in the ad is shaped collaboratively, based on who’s bidding and how much.
- The higher the bid, the more the language bends toward that brand’s tone.
- No one advertiser "wins" the space; instead, they co-author the content with the AI.
The framework considers how costs are calculated, too. Brands pay only when their input shifts what the AI would have written on its own — not for visibility, but for influence. That shift is measured by something called Total Variation Distance (TVD). It’s a statistical way of measuring how much a brand’s model changed the default output.
So instead of tracking impressions or click-through rates, advertisers would get:
- Token-level heatmaps
- Real-time influence scores
- True cost per generative impact
From landing pages to language models
This model challenges how advertisers traditionally think about campaigns. Brands wouldn’t submit copy or design landing pages. Instead, they would fine-tune their LLMs that capture how their brand should speak. That model then becomes their bid engine.
The campaign manager sets the goal, adds a budget, and the AI does the rest. Every time a relevant user query triggers the generative response, the auction evaluates:
- Which advertisers have models relevant to this topic?
- Which have active bids?
- Whose influence should shape each word of the response?
The bigger shift: from Search Engine Optimization to Generative Engine Optimization
This ad model shows that we’re moving from retrieval-based visibility (like SEO) to generation-based visibility. That means marketers now need to optimize how and when an AI talks about their brand.
Google is already dipping its toes into this with AI Overviews. The company has launched ads in AI Overviews. These ads appear inside generative responses, integrating into the context of what the AI writes. Google is also testing ads in AI Mode.
Meta is also heading in this direction. The company announced plans to roll out fully automated ad campaigns by the end of 2026. Meta is looking at a future where businesses don’t write ads at all. Its system will handle all the processes based on business goals and product feeds.