How should ads behave in AI chat & search?
It’s 2025 and chat interfaces are everywhere. Products like ChatGPT and Claude are seeing record use. But still… no ads. They are seemingly inevitable, so how might they look when they arrive?
Opportunity
Let’s think about surfaces like OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Search AI mode. These are generalist AI products in the sense that they use LLMs to target a broad spectrum of topics. The user describes their intent (in some combination of text, audio, or image) and sees a set of results (in the form of paragraphs, web links, or other summarized content) from the model. While the format and modality is new, it resembles the search problems that we've had for the last 25 years.
But that format change is meaningful for advertising. It’s important to think about how these ads will be different from those we see on search today.
Today’s search ads are carefully vetted by marketing people who write the copy and upload the creative assets to an ad manager. Then they buy a set of keywords or user attributes that they think will fit. The ad remains static no matter what kind of intent or place in the funnel a particular user might be. To target multiple sets of users or other attributes, you create and place more variants of your ad.
With AI in search, we not only have the opportunity for a number of different ad formats and placements, but also an opportunity to change how users interact with ads.
The AI ad
It seems that the medium of AI and its ability of just-in-time generation points to the fact that AI ads should be AI-generated. (I would be kind of bummed if this weren't the case.) Traditional web search has always had a gnarly problem of users providing a two or three word search query; it makes query understanding challenging and the retrieved results less relevant.
AI search users, however, seem willing to give far more (and more specific!) information than a traditional search session. So the model has more to go with and can perhaps discern a user looking to buy a camping tent from one who is looking to merely go on a hike.
An ad in the context of a chat
Maybe a particular merchant who sells outdoor gear would be interested in both of those users. Traditionally, they might place an ad for the first person showing a specific tent and another ad for the second person that does more to generate awareness for their brand of outdoor gear. But now they can buy “outdoor queries” as a whole and target any user that falls within that realm with a just-in-time generated ad tailored to the session.
In the same way that users can express their intent at a higher level, advertising buyers can express their target intents at a higher level, too. They can dump a product catalog and a folder of images from their latest photo shoot into the prompt and trust that the model will generate advertisements grounded in those facts that speak to the individual at the time of their search.
(Here’s more on declarative apps and new workflows with AI.)
Trust
The most common place users will see an ad like this is inline, within the chat. As the model incants its particular response, they will just appear in line with the response or to the side with the other corroborating information.
It should have a bit that identifies them as such. This continues the tradition of identifying what you’re seeing as an ad, which has a direct impact on user trust. There are a couple different ways we can accomplish this, however, and some are going to play into the new format that chat provides.
Ads should contain human-understandable methods for explaining why they are shown.
As the stakes for trust increase, so should the ability for these ads to be explainable. They should transparently communicate the rationale for why they were chosen and why they are showing for this particular query. Doing so should be done with care; a good method will feel less like satisfying a legal requirement and more about showcasing the ad’s utility.
Interaction
Using the metaphor of human conversation, an ad here is like someone sidling up to a discussion between two people at a party. This can land anywhere between a welcome contribution or a bother – it all depends on the context.
If a user decides to interact we should make it a great experience. The user may wonder, does this dress come in my size, and we can answer that right then and there.
An ad can open a "companion" web page to quickly serve information.
Or I might want to converse directly: are there [paper source] locations on the way home from work? There is a natural seam in the conversation here, and maybe we begin a thread, so the user can resume their main conversation if this doesn't work out. Or we could pop it into its own view, and reinforce the idea of a private conversation sandboxed from what I was just talking about. A brief aside from the party conversation while you buy an anniversary card.
Another thought: brands have long sought ways to engage (eye roll) with individuals. So they should embrace this opportunity for interaction by creating programs/gems/etc that know everything about the brand. Huge context windows and RAG offer the best chat customer experience yet – and they are finally meeting the user at a time when chat makes sense. So planning a trip becomes an easy segue into a hotel booking. The ad is just one output from this program.
New formats
There’s more fun to be had with the new formats you can make.
Prompt and refinement suggestions seems like an easy one. It could even be helpful to the user if you found a way to make it less on-the-nose and more aligned with sense-making queries.
Sponsored prompt suggestions
It’s probably better to lean into utility over gimmicks here. Focus on the stuff for which you can guarantee a great user experience.
Ads can take the form of just-in-time coded apps that serve the specific intents of the conversation.
AI can easily build ephemeral apps, and we’re getting used to seeing them in the chat surface. the experience can build itself in the moments before its needed and respond to the specific needs of the conversation. One way to think about this might be one ad ≈ one tab on your brand’s app or one page on its website. Small chunks that can pay off for the user.