AI chat is starting to look more and more like a media channel, which means someone’s gotta figure out the ad model.
Although chatbot conversations are loaded with commercial intent – ChatGPT knows it – there’s still no clear playbook for monetizing those interactions without messing with the user experience.
Search advertising specialist adMarketplace is one of the companies experimenting with what performance advertising could look like in that environment.
On Tuesday, adMarketplace announced the beta launch of AI Discover, its ad product for inserting sponsored placements into AI chat experiences, an expansion of a six-month alpha program with Aria, the Opera browser’s native AI assistant.
During the alpha, adMarketplace ran more than 2 million sponsored placements across roughly 700,00 AI queries, giving it an early look at how often people’s chats contain something that can be turned into an ad opportunity.
“The majority of what we’re seeing falls into some form of a discovery category where a consumer is trying to learn something, to figure something out,” said Sam Cox, adMarketplace’s chief product officer. “It’s a treasure trove of information.”
Search now, buy now
That early run convinced adMarketplace that people are willing to shop within AI chats.
The beta is about seeing how that behavior plays out on other AI surfaces, including Dupe, an AI search engine for affordable product lookalikes, and buy-now, pay-later platform Sezzle, which is building a chat experience of its own.
Under the hood, AI Discover runs on two adMarketplace systems. One is called Commercial Intent Vector, which scans a chat for commercially relevant topics like products, brands and categories and turns them into ad targeting signals. The other is Arena, an ad-decisioning engine that takes those signals and decides which sponsored placements to show.
Arena matches the commercial intent output against an index of more than 200 million product ads and ranks potential sponsored placements based on relevance.
But from a user’s point of view, the ad experience just looks like extra options inside the chat. Ads can show up as text-link ads, hover-over ads or sponsored results that sit alongside or just below the model’s answer.
During the alpha, a typical conversation surfaced roughly three possible sponsored placements per consumer interaction. In some cases, though, a single answer would surface dozens of monetizable opportunities.
“We’ve seen up to 50 ad slots come back,” Cox said. “In the old world, if you weren’t in the top three ad slots, you were toast, and if you weren’t above the fold, you weren’t getting a click, but we’re starting to see different behavior from within LLMs.”
From query to cart
Cox knows of what he speaks from firsthand experience.
One of his hobbies is fixing old cars, and he recently used AI search to figure out which components he’d need to keep one of them running.
“It gave me a list of four and I clicked on all four and bought all four,” Cox said. “Then I asked what else was going to go wrong next and it gave me 10 more parts – and I literally bought all 10.”
Rather than bouncing between multiple forums, search results, tabs and ecommerce sites, Cox made 14 purchases during the space of a single conversational thread. “We’re seeing a compression of the funnel,” he said, “and that’s a huge change in user behavior.”
But Cox is adamant that the ad stack doesn’t influence a model’s results. If people are going to trust AI enough to shop inside of chat, he said, the answers and recommendations they see can’t be influenced by ad dollars.
That’s why adMarketplace doesn’t receive full prompts from its LLM partners and stresses that its commercial intent tech runs inside of a partner’s environment, not on its own servers.
“People believe this to be a private experience and that’s an important thing for user trust,” Cox said. “Consumers don’t want the ad model impacting the reasoning model.”
Clicking into place
But building consumer trust is only one part of the puzzle. The other question is how to charge for AI search ads.
All of adMarketplace’s AI Discover inventory runs on a cost-per-click model, not on impression-based pricing. Advertisers only pay when someone clicks from a chat to their site, which makes sense for a new channel where a single answer can surface multiple relevant options in short order, Cox said.
“A CPC model is good for advertisers,” he said, “because they’re not paying $60 or $100 CPMs just for exposure without any relevance to what’s in the actual chat.”
The AI Discover beta will run from June through December in the US, UK, Germany and France with roughly 10 advertisers and five LLM partners, including Opera AI, Dupe and Sezzle. Cox declined to share whether adMarketplace is having conversations with any of the biggies, such as OpenAI.
But, hey, it’s still early days. For now, AI Discover is less a finished product than an ongoing experiment in how people actually use AI search and how their habits are shifting.
“We’re in a process of discovery here as well,” Cox said.
