Google’s Token Auction: When LLMs Write Ads in Real Time

The world of PPC (Pay-Per-Click) advertising is on the verge of one of its most profound transformations. Until now, advertisers have competed for advertising slots on search results pages, where the platform simply displayed ads they had written in advance.

However, a new generation of large language models (LLMs) offers a radically different alternative: ads that are not selected, but written in real time through an auction based on the words—or tokens—being generated. This model is known as a token auction.

The idea may sound like science fiction, but it is now supported by real research. Google Research and the University of Chicago have published a paper outlining a theoretical and practical framework for how advertising auctions could operate in the age of generative AI.

The concept of embedding advertising within AI-generated content is not entirely new. As early as 2018, Google patented a technology called “Using Different AI Entities as Advertising Mediums,” which explored how virtual assistants and chatbots could integrate sponsored messages into conversations.

With today’s advanced LLMs, however, that early vision becomes far more dynamic. It is no longer simply about inserting ads into a conversation—the advertisement itself becomes part of the conversation.

Token Auctions: A Generative Advertising Model

In traditional Google Ads, advertisers bid on keywords. The system then selects an appropriate, prewritten advertisement based on factors such as bid amount and quality. The advertising creative remains fixed, while the auction determines where it will be placed.

In the token auction model proposed by the researchers, this paradigm changes. Advertisers no longer bid for advertising placements; instead, they bid to influence the words generated by an LLM.

According to the research, the process would work as follows:

• Each advertiser submits a single bid.
• Along with the bid, the advertiser provides a language model representing its brand voice, tone, and messaging preferences.
• The system generates the response token by token—or word by word—while taking each advertiser’s model into account according to the value of its bid.
• Instead of selecting a single winner, the system combines the influence of multiple advertisers within the final output. The higher the bid, the more strongly the generated language shifts toward that advertiser’s preferred voice.

To combine the competing language models, the researchers examine two strategies:

• Linear aggregation: A weighted average that preserves incentive compatibility and responsiveness to bids.
• Log-linear aggregation: A more complex method that may violate incentive compatibility under certain conditions.

Understanding this distinction is critical. Advertisers who do not know which aggregation model is being used could overspend in exchange for minimal influence—a costly mistake in this new type of auction economy.

Redefining Brand Preferences

Under this framework, advertisers would no longer submit static ad copy or landing pages. Instead, they would train an LLM to communicate in their brand’s voice.

The model would act as a dynamic representation of what the brand might say in any given context.

The challenge would no longer be to create one outstanding headline. Instead, advertisers would need to engineer a probabilistic system capable of reliably generating brand-aligned language within real-time conversations.

Protecting Privacy and Reducing Friction

Another important aspect of the proposed system is its technical decoupling.

The primary generative model—for example, the model responding directly to the user—would not have direct access to the internal logic of each advertiser’s LLM. Instead, every advertiser would privately calculate the probabilities of the next possible tokens and submit those probabilities to the auction system.

This means that brands could participate without exposing their proprietary models or internal logic, while the central generation engine remains efficient and modular.

The Business Model: Paying for Actual Influence

In a traditional Vickrey auction, an advertiser pays only when its bid changes the outcome. The same principle would apply here: an advertiser would pay only when its influence causes the system to generate a token that differs from the token that would otherwise have been selected.

The token auction model is built around this principle.

How Would This Influence Be Measured?

The proposed measurement is a statistical metric known as Total Variation Distance (TVD). It measures how far the final output shifts from the default result because of a particular advertiser’s input.

This represents a move away from clicks and impressions toward token-level ROI. For the first time, it may become possible to measure a brand’s influence at the level of individual words.

Simulation: Does the Model Actually Work?

The researchers tested the model using Gemma 7B, an open-source large language model, and two hypothetical advertisers: one with a formal communication style and another with a more casual tone.

Each advertiser expressed different stylistic preferences through its own model. The results showed a clear correlation between higher bids and stronger influence over the tone and wording of the generated response.

The test question was: “What is a good activity for the weekend?”

The charts and tables included in the research demonstrate how this influence could be modeled, tracked, and monetized.

From SEO to Generative Visibility

This model is not only about advertising. It reflects a broader transition from retrieval-based systems, such as traditional search and SEO, toward generation-based visibility—also known as GEO, or Generative Engine Optimization.

In SEO, you optimize a web page so that it appears in search results.

In GEO, you optimize your language and information so that they appear as part of a generated answer.

Token auctions suggest that we may soon see the emergence of paid GEO, where the answer itself is shaped in real time by brands that pay to influence its composition.

What Could the Advertising System of the Future Look Like?

It is difficult to predict exactly how such a system would operate, but one possible workflow could look like this:

• A brand fine-tunes its own LLM to reflect its style, tone, and communication principles.
• A campaign manager defines specific objectives—for example, “Mention our brand when someone asks about a romantic holiday”—and submits a bid.
• When a user’s question triggers a generative response, the token auction evaluates, token by token, which advertisers have relevant models and active bids.
• The system collaboratively generates the final response, influenced by the brands participating in shaping its language.
• The advertiser dashboard no longer focuses primarily on clicks. Instead, it displays token-level heat maps, influence scores, and the actual value or cost associated with each generated impression.

What Does This Mean for Marketers?

This is no longer about selecting which of five ad variations performs best. It is about influencing what the AI will say.

• Static creatives disappear: The system generates advertising messages in real time.
• Language engineering becomes more important than traditional copywriting: Success depends on probabilistic language modeling rather than catchy slogans alone.
• Multi-brand answers become standard: Several advertisers may influence or appear within the same generated response.
• Presence replaces placement: Brands aim to shape the message itself rather than merely appear beside it.
• A new ROI model emerges: Performance is no longer measured only through clicks or impressions, but through the advertiser’s influence on the final output.

Google’s AI Advertising Today

This is not purely theoretical. In May 2024, Google began testing Search and Shopping ads within AI Overviews and AI Mode.

These sponsored messages now appear within generative results on both mobile and desktop devices.

According to The Verge, more advertisements are now being embedded in AI-generated answers than ever before.

For now, these ads still use a traditional format: they are labeled, clickable, and visually distinguishable from the surrounding content. However, the direction is clear—advertising is moving directly into the content experience.

The Future of Paid Advertising

This transformation is not limited to Google. Meta has announced that, by the end of 2026, it aims to introduce fully automated paid advertising campaigns in which businesses may no longer need to write their own ads.

An advertiser would simply define an objective—for example, “Sell green running shoes”—and the system would manage the creative, targeting, testing, and optimization automatically.

Taken together, the developments we are seeing from Google and Meta point toward a future in which paid media is no longer only about targeting or placement.

It is increasingly about collaborating with machines that generate, optimize, and deliver brand messages in real time.

We are moving:

• From prewritten advertisements to AI-generated responses.
• From manual optimization to real-time probabilistic influence.
• From click-through rate (CTR) to token-level brand presence.

Whether this future is shaped through Google’s token auctions or Meta’s fully automated campaigns, the common direction is clear: paid advertising is becoming generative.

For marketers, moving forward will require new skills, new strategies, and a deeper understanding of how to influence AI-generated outputs—potentially without ever writing a traditional advertisement.

Source: Search Engine Land