Every few years, a new platform launches and brands that normally take six months to approve a creative brief suddenly find urgent budget. We’ve seen it with TikTok, with influencer, with connected TV. Now it’s happening with OpenAI ads. We shared our thinking on this with The Information recently, and it’s a conversation worth continuing here.
Clients who have been cautious about testing proven formats are ready to move fast on something that, by most measures, has more unknowns than anything that’s come before it. The motivation is understandable. Search demand has been softening for a while, and OpenAI feels like the place where that intent is going. But the instinct to treat it as a like-for-like replacement is worth examining before any budgets move.
The fear driving the rush
A lot of the urgency we’re seeing from clients right now comes down to search anxiety. They’ve watched their organic and paid search volumes flatten or decline, they’ve heard about AI overviews eating into click-through rates, and now there’s a platform where people are clearly going to ask questions and make decisions. The logic writes itself: get in front of them there.
The problem with that framing is that it assumes the job is to replicate search, just on a new surface. That however sells the opportunity short and also sets up unrealistic expectations. Intent in a chat environment is real, but it’s not the same as a search query. The context is different, the format is different, and the relationship between the user and the platform is different in ways that matter for how ads will actually perform.
Why this surface is genuinely different
One thing that separates ChatGPT from Meta or Google, at least emotionally, is how personal the experience feels. People use it to think through decisions, work through problems, have conversations they might not have with other tools. That one-to-one quality is not incidental. It’s the product. And it creates a dynamic that brands advertising there will need to think through carefully.
On a more practical level, the ad environment is non-deterministic. The same question can produce a different answer depending on a thousand variables, and there are limited tools available right now to control the context in which an ad appears. Brand safety, which many advertisers take seriously and set strict parameters around, is harder to guarantee here than on almost any other platform. An ad might appear alongside content that’s topically relevant but tonally off, and the chatbot’s personality and tone in that moment will shape how the ad lands. That’s a layer of uncertainty that doesn’t exist in display or search.
Three things worth working through before committing budget
The CPM math doesn’t hold up yet. At its core, this is a contextually targeted native format. That’s not a new media type, and the industry has been buying similar placements for years, often at a fraction of what OpenAI is currently pricing. When comparable inventory is available at significantly lower cost, the question for any performance advertiser is: what else are we getting for the premium? Right now, that’s genuinely hard to answer.
Measurement is the other major blocker. Without the tools to properly attribute performance, validating the investment becomes very difficult. The minimum viable approach is incrementality testing with holdouts: splitting audiences randomly, holding back a control audience, running it for enough time to get significant data, and measuring how rates differ between the two groups. That’s how you get a defensible read on what the ads are doing. It’s not the same as the kind of attribution most clients expect before scaling spend, but it’s the most honest signal available right now.
Audience clarity matters too, and you don’t know whose intent you’re buying. “High-intent users” is a compelling pitch, but intent is only useful if you know whose intent it is. ChatGPT’s user base is not a monolith, and it’s also not the entire internet. There’s already meaningful fragmentation across LLM platforms, and the demographics likely vary significantly between ChatGPT, Gemini, and Claude. Assuming you’re reaching a broad, high-value audience without that data is a leap.
On how to test well, if you’re going to test
The framing that makes the most sense right now is treating this as an experiment with a clear hypothesis, not an activation with expected returns. That means going in with specific questions you want answered: which verticals perform, where in the funnel this format fits, what the CPM justification looks like against comparable placements. It also means not letting internal structures get in the way. Whether this sits in a search budget or a programmatic budget is less important than making sure the right people are across it and the test is set up to actually produce learnings.
The bigger picture on LLM fragmentation
One thing worth keeping in mind as OpenAI continues to build out its ad product: it does not have the market position that Google and Meta had when they launched their ad platforms. There are real, well-resourced competitors with growing user bases, and the audience is already splitting across platforms. The total addressable audience on any single LLM is a subset of what it might appear to be, and that fragmentation is going to continue.
That also means the ad campaigns these platforms run to attract and retain users are going to become increasingly interesting to watch. How OpenAI, Google, and Anthropic each position themselves to different audiences will tell advertisers a lot about where specific audiences are actually going.
The right posture right now
There’s a version of this that goes well for brands, and it looks like clear hypotheses, realistic measurement expectations, and genuine willingness to learn from what the data shows. It doesn’t look like redirecting search budgets wholesale in the hope of recapturing lost volume, or committing to a platform before understanding who is actually on it.
The opportunity is real. The timeline for it becoming a proven, scalable channel is not as short as the current momentum might suggest. Going in with that expectation, rather than the one being sold, is probably the better bet.
