Sometime this week, a media buyer at a US brand bought an ad inside ChatGPT. Not a banner on a website about chatbots. An actual placement beneath a ChatGPT answer, purchased through Amazon DSP, the same console that buyer already uses for display and streaming video spend.
Search Engine Land reported the pilot on Thursday. What it forces every performance marketer to confront: the “results page” you’re buying your way into is now a single conversational answer, and the ad sits underneath it. If you spent a decade learning Google’s auction, some of that knowledge transfers. A surprising amount doesn’t.
What actually shipped
The mechanics are simple on the surface. Amazon advertisers in the US pilot can now buy ChatGPT ad placements through Amazon DSP. The units appear as labeled advertisements beneath ChatGPT’s answer, sold the old-fashioned way: by the click (CPC) or by the thousand impressions (CPM).
Two details matter more than the format. First, OpenAI decides placement. Amazon brings the advertisers and the pipes; OpenAI keeps editorial control over what appears next to its answers. Second, this is the first programmatic on-ramp into a surface with more than a billion users. Whatever norms get established in this pilot will shape every AI-answer ad product that follows.
It was also part of a crowded day. OpenAI shipped five things on Thursday: the Agents API, ChatGPT for Financial Services, the GPT-Live-1 voice model, this Amazon ad pilot, and a 27-month government agreement. The ads pilot got the least attention and may matter most to marketers.
Why your search instincts will misfire here
On Google you compete against ten blue links and a wall of shopping boxes. Attention splits. On ChatGPT the user reads one answer, and your ad arrives after the model has already responded. That changes the ad’s job completely.
In search, an ad interrupts the task. Here, it follows a finished answer. Sometimes the answer ends the job (“what’s the cancellation policy for this airline”) and the ad is decoration. Sometimes it leaves the user wanting a next step (“best CRM for a five-person agency”) and the ad is the bridge. You won’t always know which situation you bought.
There’s also the auction question. Google publishes reams of documentation on quality score, ad rank, and bidding mechanics. OpenAI has said nothing comparable. Placement gets decided unilaterally, and early buyers are flying without instruments. Budget accordingly.
The measurement problem nobody has solved
What counts as a conversion when the answer itself might complete the task?
If a user asks ChatGPT for the best stroller for apartment stairs, gets a thorough answer naming three brands, and then clicks your ad to buy, did the ad convert them or did the answer? If they never click because the answer satisfied them, was that a lost sale or a free brand impression? Nobody has clean answers yet. Set that expectation with your team before the first dollar goes out. Pilot data will be muddy. That’s normal.
Should you get in now?
My read: if you already buy through Amazon DSP and you sell to consumers, yes, with a small budget and low expectations. Learning a surface early is cheap. The brands that figured out Amazon search ads before the case studies existed built durable advantages over the ones that waited.
Three checks before committing spend:
- Pull your analytics. If chatgpt.com and other AI referrers already send you traffic, the surface is proven for your category. If they send nothing, an ad won’t fix that on its own.
- Ask your Amazon rep about pilot scope, minimums, and reporting. The details are still account-by-account.
- Write your ad for the position it occupies: the next step after an answer. A search-style keyword hook will read wrong in this slot.
I’d hold off if you’re B2B with a six-month sales cycle or in a regulated category. The measurement infrastructure isn’t there yet, and a muddy pilot won’t tell you anything your dashboard can use.
Three questions that will set the market
Watch these. The answers will set the floor for the entire AI-ads business.
One: do users tolerate ads in what they think of as an assistant? Chat still feels like a tool, not a media channel, to most people. Labeled units beneath answers are the gentlest possible test of that boundary. Early numbers, when they leak, will be the most quoted figures in this space.
Two: does OpenAI ever publish placement rules? Unilateral control is workable in a pilot and unacceptable at scale. No advertiser will build a serious budget on a black box.
Three: what happens to the organic answer? If the model names brands in the answer and sells ads beneath that same answer, every content and SEO team needs to watch how the two interact.
What to do this week
Keep it small and concrete. Pull AI-referrer traffic from your analytics. Ask your Amazon DSP rep whether your account is in the pilot’s scope. Draft two or three units written as the step after an answer. Set a test budget you’re comfortable spending for information rather than conversions, and decide in advance what result would make you scale up.
The worst move is treating this as search wearing a new skin. It’s a new surface with different physics, and the pilots running right now are the cheapest tuition you’ll ever get on it.
