Amazon Put Its Advertisers Inside ChatGPT

On September 10, Amazon announced that select US brands advertising through Amazon Ads, including its demand-side platform, can run campaigns inside ChatGPT. Delta Vacations is among the first named. Amazon sets up and manages the campaigns through its DSP; OpenAI's own ad system handles all delivery and placement decisions. Units appear as labeled text or images beneath ChatGPT responses, bought on CPC or CPM.

The context number: OpenAI said last week that ChatGPT Ads reached a $1 billion annualized run rate in under 200 days, with tens of thousands of advertisers. ChatGPT had 900 million weekly active users as of February.

Two hundred days to a billion-dollar run rate is the fastest ramp of any ad product in history, and it is worth being precise about why.

The demand was already there

New ad platforms usually fail at the supply side of the marketplace — plenty of inventory, no advertisers who trust it. ChatGPT skipped that problem entirely, because the advertiser demand for intent-adjacent placement has been starved for two years.

Search advertising works because someone typing a query has declared intent. That is the whole mechanic behind a $200 billion business. What has been happening since 2023 is that a growing share of high-intent queries — which laptop should I buy, where should we go in March, what's the best CRM for a 12-person team — migrated from a search box into a chat window, and the advertising did not follow.

Every performance marketer has watched that migration and had nowhere to spend against it. The $1 billion run rate is not OpenAI building demand. It is OpenAI opening a valve on demand that had accumulated.

Why Amazon and not just OpenAI's own sales team

The division of labor in this deal is the tell. Amazon handles campaign setup and management. OpenAI handles delivery and placement.

Amazon brings three things OpenAI does not have. Advertiser relationships — tens of thousands of brands already running budgets through Amazon DSP, with trafficking teams, creative pipelines, and measurement habits built around it. Purchase data — Amazon knows what people actually bought, which is the closed-loop signal that makes conversion attribution possible. And buying infrastructure — a DSP that already handles targeting, frequency, pacing, and reporting at scale.

OpenAI keeps the part that is genuinely its own: deciding what appears where, in a surface where relevance is judged against a conversation rather than a keyword.

For Amazon, this extends a pattern that has been running all year. Amazon has secured ad inventory access across Netflix, Roku, Spotify, SiriusXM, Disney, Hulu, and ESPN. The strategy is consistent and unglamorous: be the buying layer for premium inventory you do not own. ChatGPT is the highest-growth inventory on that list.

The part advertisers should be nervous about

OpenAI's ad system handles all delivery decisions. Amazon manages the campaign; it does not control placement.

That is a meaningful loss of control relative to how DSP buying normally works. In conventional programmatic, the buy side bids on specific impressions with known context. Here, the buy side expresses intent and budget, and a model decides which conversations get the ad. The advertiser does not know — and structurally cannot know in advance — what the user asked to trigger the placement.

Brand safety in that model is not a blocklist problem. It is a model behavior problem. A hotel ad surfacing under a travel question is the happy path. The same ad surfacing under a question about a family emergency abroad is the one that generates a screenshot.

The units are labeled, which handles the disclosure question but not the adjacency question. Advertisers accustomed to controlling adjacency are being asked to trust a system they cannot audit. At a $1 billion run rate, enough of them are accepting that trade to make it a real business.

What this does to commerce workflows

Three practical consequences for anyone selling online.

Product feed quality becomes an AI-visibility problem. The pilot generates product ads from catalog data. Whatever your feed says about your product is what the model has to work with when deciding whether your ad answers someone's question. Thin, inconsistent, or poorly attributed catalog data was already a conversion problem; it is now a placement eligibility problem.

Attribution gets harder before it gets better. A conversational placement sits somewhere between search intent and display discovery, and neither existing attribution model fits it cleanly. Expect a period where the reported ROAS on this inventory is unreliable in both directions — and expect Amazon's purchase data to be the thing that eventually resolves it, which is precisely why Amazon is in the deal.

The organic side just got more competitive. Every brand that has been optimizing for appearing in AI answers now has a paid competitor for the same real estate. The work of being the answer a model reaches for has not stopped mattering — it just stopped being the only path onto the page.

The number to watch

Not the run rate. Retention.

A billion-dollar annualized run rate at 200 days is built substantially on advertisers testing a new surface. Test budgets are easy to win and easy to lose. The question that determines whether this is a real ad platform or a very fast experiment is whether those advertisers are still there in Q2 2027 at higher spend, with measured performance that justifies it.

Amazon's involvement makes that more likely, because Amazon's DSP customers do not run tests they cannot measure. That is the actual value of this partnership — not the inventory access, but the measurement discipline that comes with it.