Amazon Owns the Buyer, OpenAI Owns the Moment

Until this week, Amazon treated ChatGPT as a rival to keep out. It barred OpenAI's assistant from crawling its retail catalog while it built its own shopping AI — and, at the same time, quietly became the top retail advertiser buying space inside ChatGPT through the spring and summer.

On September 10, the posture flipped. Amazon Ads is now selling that space to everyone else.

Select U.S. advertisers can extend campaigns into ChatGPT through Amazon DSP, with Delta Vacations among the first to test it. The deal hands OpenAI an entire demand operation — advertiser relationships, campaign management, commercial signals — while OpenAI keeps, for itself, the only decision that finally matters: which ad runs, when it appears, and who wins the moment.

That split is the whole story. It's also where the unanswered questions live.

How the buying works

Amazon is offering ChatGPT Ads through Amazon DSP as a managed service — its team helps set up and optimize campaigns rather than opening the inventory to self-service trading. The U.S. pilot supports cost-per-click and cost-per-thousand buying. Ads surface below ChatGPT's responses as labeled text and image units, and Amazon is also running product-feed campaigns that turn a catalog into ad assets automatically, so retailers can expose a wide range of products without building each unit by hand.

From the buyer's side, it looks like an ordinary DSP extension. A brand already working with Amazon sets a budget, hands over assets, picks targeting, and gets reporting through an existing relationship.

Behind that interface, the delivery environment is nothing like a website, an app, or a streaming feed.

A publisher sells inventory attached to a page, an audience, a content category — things you can name before the auction. ChatGPT sells a moment inside a conversation that is still moving. OpenAI's system reads the current exchange, the landing page, the creative, the advertiser's context hints, and, where personalization is on, selected signals from the user's broader ChatGPT history. Those hints aren't keywords, and they don't guarantee placement; eligibility and ranking run through a relevance-weighted, second-price auction where a sharper, more relevant ad can beat a bigger bid.

So Amazon DSP is the on-ramp, not the destination. ChatGPT inventory has entered the programmatic supply chain without becoming interchangeable with anything already in it.

Amazon owns the buyer, OpenAI owns the moment

Amazon's contribution is demand. It has the brand and agency relationships, a large campaign-management operation, and shopping, browsing, and streaming signals that shape media strategy — the machinery of a company that neared $70 billion in ad revenue last year and runs the third-largest digital ad platform in the market. Its DSP already places ads across Amazon properties and third-party inventory, including for advertisers that sell nothing on Amazon at all.

OpenAI owns the room. Its systems decide whether a given conversation should carry an ad, whether a campaign fits the exchange, and where the unit lands beneath the answer. OpenAI writes the ad policies and the user controls, too.

That produces a layered operating model, and a layered accountability problem.

Amazon can recommend settings and optimize spend; OpenAI runs the auction and makes the final call. When a campaign underdelivers, drifts into a context nobody intended, or produces a disputed result, responsibility is split across two companies and two systems — and the public materials don't explain how optimization instructions pass between them, or whether Amazon sees enough delivery data to diagnose a bad campaign on its own rather than waiting on OpenAI's aggregates.

Agencies will want a responsibility map before they commit real budget. The party taking the money, the party choosing the ad, and the party reporting the result may each see a different slice of the same transaction.

Familiar metrics, unfamiliar denominator

Pilot advertisers get reporting they recognize: impressions, clicks, cost per result, CPM, CPC. OpenAI's own Ads Manager adds spend, CTR, average CPC and CPM, and conversions, with tracking parameters, an OpenAI Pixel, a Conversions API, and outside partners such as LiveRamp, which can pass online or offline conversions server-to-server.

The numbers make ChatGPT legible inside a media plan. They don't tell you what it's worth.

The Amazon announcement doesn't say which of those measurement paths run through the managed service, or name an attribution window, an invalid-traffic standard, a viewability definition, or the level at which data can be exported — or whether Amazon's numbers and OpenAI's will reconcile without a fight.

And a click earned under a long travel-planning exchange is not a click on a search ad. That difference can be the whole value of the placement — or it can flatter it. A high conversion rate may just mean the user had already decided before the ad appeared. Incrementality doesn't fall out of a good CPC. Agencies that want to know whether the ad changed behavior, rather than sat near a decision already made, will have to build that test themselves — clean conversion instrumentation, controlled geographies, exposed and unexposed groups — before the first dollar goes in.

The privacy line has an asterisk

OpenAI's promise is specific: advertisers don't get your chats, chat history, memories, or personal details. They receive aggregated, non-identifying performance data. Users can turn off personalization, clear ads data, hide a placement, and check why an ad appeared.

The asterisk is that turning off personalization doesn't turn off contextual targeting.

The current conversation still feeds ad selection; with personalization on, past chats, memory, and prior ad interactions feed it too. The boundary OpenAI draws is between using what you say and disclosing it — an advertiser can benefit from an inferred intent without ever seeing the words that produced it.

Amazon adds a second data environment. It sells its DSP on deep shopping, browsing, and streaming signals, and Delta Vacations has said it will use consumer insights to shape when and how it shows up in ChatGPT. Neither company explains the data flow behind that sentence.

Do Amazon's signals build audiences before anything reaches OpenAI? Do they become targeting selections or context hints? Do they only guide Amazon's managed-service advice? Is there any matching across the two systems, and what stops a commercial profile from being fused with conversational intent? "Advertisers can't read your chats" answers one serious question and leaves inference, matching, retention, and optimization untouched. Agencies will need a data-flow diagram that names what enters each platform, what crosses between them, and what comes back in reporting.

Placement suitability is now a moderation problem

OpenAI keeps ads out of sensitive and regulated territory and away from accounts it identifies as under 18. On a conversational surface, enforcing that is harder than it sounds.

A publisher can tag a page as travel or finance or sports before the auction runs. A chatbot has to read an exchange that can change direction in a sentence, disclose something personal without warning, or slide into a sensitive subject mid-thread.

Suitability stops being a static taxonomy and becomes a live interpretation the platform has to keep updating as the conversation moves.

Picture the pilot's own example. A user asks for vacation ideas; the ad is a clean fit. Three messages later, the same user mentions the trip is for a dying parent, or that they've just lost the job that was paying for it. The context that made the placement appropriate is gone, and the advertiser never knew any of it happened.

They can't audit it, either, because OpenAI won't show them the exchange — and shouldn't. Protecting the user's privacy is exactly what limits the advertiser's ability to verify suitability. Both goals are legitimate; the tension between them is real and currently unaddressed. Independent testing, incident reporting, and disclosed suitability categories could give advertisers assurance without exposing a word of anyone's conversation. The pilot materials don't describe anything like that yet.

The ad is stapled to the answer

OpenAI promises that ads don't influence answers. It makes no promise about the reverse.

Your sponsored unit can sit directly beneath an answer that pans your product, points the reader to a competitor, or says something about you that's flatly wrong. On the screen, the ad and the answer are one moment — one implied endorsement. Answer independence protects OpenAI. It does nothing for the advertiser stapled to the paragraph above.

Creative approval won't manage that. Before the first impression runs, a media plan needs scenarios for the conflicting answer, the sensitive turn, the misleading proximity, the user complaint — and a contract that names who investigates a bad placement, who validates delivery, who owns discrepancies, and what each party can actually see when something goes wrong. Report ChatGPT as its own experimental line, not blended into search or display, until those answers exist.

What the pilot hasn't settled

Amazon has made ChatGPT ads easy to buy. It has not made them easy to answer for. The interface now looks like every other line in a media plan — same objectives, same CPC and CPM, same managed-service hand-holding — while the thing being bought still behaves like nothing else on that plan.

A brand can win a placement it never sees, beneath an answer it can't read, in a conversation that turned toward a crisis three messages earlier, chosen by an auction it doesn't run and reconciled by a partner that may not see the whole transaction either.

Conversational advertising may well become a real performance channel. Whether it does will depend on things this pilot hasn't settled: whether inferred intent can be made useful without eroding privacy, whether agencies can measure incremental influence instead of proximity, and whether responsibility stays legible when the buying system and the delivery system belong to different companies.

Familiarity is arriving faster than accountability. Buy the test. Refuse the assumption that the two are the same thing.