Advertising in the Age of AI Conversations

AI chatbots are upending the old rules of advertising. What used to be a game of bidding on keywords to secure a top spot on a results page is becoming a race to weave your ad into organic conversations.

But when should a chatbot introduce an ad? And how can it do so without feeling intrusive or disrupting the natural conversation flow—a feature that makes large language models (LLMs) valuable in the first place?

A series of new papers by Professor Ali Makhdoumi of Duke University’s Fuqua School of Business, coauthored with Saeed Alaei (Google Research) and Azarakhsh Malekian (Google Research and University of Toronto) propose a systematic approach to advertising within LLMs.

One paper focuses on timing—when an LLM platform should display an ad during an ongoing conversation. The second examines how ads can be woven directly into chatbot-generated text.

As businesses try to figure out how to adapt to AI chatbots taking over search engines, this research offers a comprehensive framework that could shape how AI chatbots reimagine the legacy of traditional advertising.

“AI is obviously going to change a lot of things, but advertising is one thing in particular,” Makhdoumi said.

Chatbot conversations develop in ‘time’

In traditional searches, people entered simple queries— such as “musicals in NYC”—and advertisers bid to place “sponsored results” at the top of those searches.

But in the era of LLMs, users engage in back-and-forth conversations with chatbots, refining their questions with the ease of natural language. Someone who searched for “musicals in NYC” might keep going, asking, for example, “which of the current shows have Sunday matinees that end before 5pm?”—in case they have a late-afternoon flight.

As the conversation unfolds, the LLM learns more about the user’s preferences. This introduces a new variable that didn’t exist in traditional search: timing. In addition to deciding which ad to offer, platforms need to decide when to advertise during the chat.

“There is a time aspect in advertising for LLMs that wasn’t present with traditional systems,” said Makhdoumi.

As the platform learns more about the user, it can better decide which ad is a better match. “But you cannot wait forever,” he said, because the longer the chat goes on, the greater the risk that the user will drop out of the conversation.

How do platforms decide when the right time is to show an ad?

A rulebook for platforms

In the paper, Dynamic Learning and Optimal Advertising Mechanism for LLM Platforms, the researchers developed a "threshold-based policy" that tells the platform to keep learning as long as the expected benefit of waiting remains high, but to stop once the current value of showing an ad exceeds a certain threshold.

To establish this threshold, Makhdoumi and colleagues created a formula to find a “target score” for every round of conversation.

If the ad the platform could serve at a certain stage generates a value below the target score, the LLM keeps the conversation going. But if the potential ad hits the target, the LLM stops learning and offers the ad.

“At each round of the conversation, the platform compares the current gain from showing the ad to its assessment of future expected gains,” Makhdoumi said. “If the current gain is larger, the platform shows the ad. If not, it will wait.”

The researchers also suggested that waiting is most beneficial when advertisers are selling highly differentiated products, which increases the risk of having a poor match between the user and the ad.

Knowing all users’ preferences is practically impossible, the researchers explained. As long as the information gathered by the chatbot meets a certain threshold, the model can still perform close to optimally.

Embedding ads in conversations

But deciding when to show an ad is only one part of the challenge.

The researchers worked on a second paper—Transforming Transformers: Content Generation with Advertiser Incentives under RoS Constraints. It examines how to insert ads into chatbot conversations without disrupting the user experience.

“If you want to embed advertisements within the text, there’s a natural way to do it,” Makhdoumi said. “You change the language a little bit, but not too much, because you don’t want to stray too far from the organic response.”

The researchers introduced a “plausibility” constraint—a formula that measures how far the language can stray from its organic answer. They then set a limit on how much the platform can change the output within an unnoticeable amount.

“In a way, we are transforming the ‘transformer,’” Makhdoumi said, referring to the underlying system that converts user words into “tokens” that the machine can understand, which then are used to build answers step by step.

Their proposed solution is a “tilted distribution,” which allows the AI system to weigh advertisers’ bids to “tilt” its responses.

The platform can also control a second lever—how much advertisers are charged—to encourage advertisers to reveal their true valuations of the ad, Makhdoumi said.

From keywords to conversions

In this new framework, advertisers will bid for conversions, rather than just for keywords. 

In traditional SEO advertising, advertisers bid for keywords such as “wedding ring,” with the assumption that placing an ad alongside those searches will generate clicks and, given a certain known rate, some actual business conversions. “But this changes, with LLM advertising,” Makhdoumi said.

With chatbots, the relationship between a keyword and a conversion is less clear, he said, because as the conversation progresses, the chatbot learns the user preferences and may steer the user away from certain products or category of products. Therefore, a keyword appearing in a chatbot response does not necessarily translate into clicks or conversions and is highly dependent on the context.

“So here it makes sense for the platform to charge the advertiser for conversions, rather than for impressions,” he said. “Because here the whole conversation, not just a single keyword, influences the user's purchasing decisions.”

This approach will allow platforms to maximize ad revenue while protecting the relevance of the answers.

“You want to start with the organic response and then change it a little bit to deliver value to the advertiser, without changing the response in arbitrary ways,” Makhdoumi said.

As conversational AI becomes the intermediary between users and information, such decisions—when and how to place an ad and how to monetize—will shape the future of digital advertising.

“Research on advertising with LLMs is still very young,” he said. “What we wanted to do with these papers was highlight the new questions and the new framework for managing advertising in this new conversational space.”