Amazon DSP Now Sells ChatGPT Ads: What Ecommerce Teams Should Prep

On September 10, 2026, Amazon Ads announced a partnership with OpenAI that lets advertisers extend their Amazon Ads campaigns into ChatGPT as a conversational ad experience. Select advertisers in the U.S. are testing it now as part of a pilot, and Delta Vacations is among the first brands named in the announcement.

The mechanics, per the reporting: buying starts through managed service, then opens to self-service through Amazon's DSP. OpenAI continues to control ad delivery and placement inside ChatGPT; Amazon provides the buying and campaign-management layer. Ad Age called the tie-up Amazon becoming ChatGPT's biggest ad-tech partner, in a program that already counts more than 50 ad-tech partners.

That is the whole announcement. What matters for an ecommerce team is what it signals about where the bottleneck has moved.

The pattern across six months of ChatGPT Ads news

The sequence from our own coverage makes the shape obvious:

February 2026 — ChatGPT ads launch in a U.S. pilot, bought directly through OpenAI.
August 12 — expansion to five more markets; August 20 — expansion across 31 European markets, with conversion optimization, geo-targeting, custom audiences and a Pixel plus Conversions API.
August 31–September 1 — the program crosses a $1B annualized revenue run rate in under 200 days, and self-serve Ads Manager opens across India, Europe, the Middle East and North Africa, with CPC and outcome-optimized bidding now the majority of campaigns.
September 10 — Amazon's DSP becomes a way to buy the inventory.

Every cycle has removed a piece of access friction: more markets, then self-serve, then the buying platform you may already run. None of them has removed a single unit of creative friction. That asymmetry is the thing to plan around.

Why a DSP matters more than another market launch

A market expansion adds geography. A DSP integration changes the plumbing.

When conversational inventory is only buyable inside OpenAI's own Ads Manager or through an agency, testing it requires a new account, a new billing relationship and a new reporting surface. Through Amazon DSP it reaches the advertisers who already run Amazon campaigns — same buying team, same budget line, same dashboards, with a managed-service on-ramp while the pilot is still gated.

That lowers the cost of trying ChatGPT ads to roughly the cost of adding one line item. It also means the number of advertisers competing for the same conversations goes up the moment self-serve opens. Buy-side access is being commoditized — the same commoditization we've watched happen on the production side, where Google put generative video creation directly inside Asset Studio and Performance Max, and where frontier video models became cheap APIs anyone can call.

What you still own: the creative and the offer

Conversational inventory is pull, not push. Someone is mid-decision — comparing options, learning a category, asking what to do next — and the ad appears inside that moment rather than interrupting a scroll. OpenAI's own framing when it added CPC bidding was that a click is a meaningful signal that an ad was relevant.

That has a practical consequence: you can't rent attention with placement alone. The creative has to earn a click from a reader who was in the middle of a thought. In practice that means:

A hook that states the product and the problem in the first seconds. There is no autoplay scroll-by doing the explaining for you.
Product-first framing. The ad is served next to a comparison the user is actively making — vague brand mood videos have nothing to attach to.
One claim per variant. If you cannot say what a creative is testing in a single sentence, it is not a variant, it is a brochure.
Vertical, short, captioned creative. The same format that already works in feed-based placements is the natural shape for a chat context, and it is the format a UGC-style pipeline produces by default.

Plan for a split measurement story

The structural detail worth internalizing: Amazon manages the buying relationship, OpenAI decides how and where ads are served. That is a two-vendor delivery chain, and attribution will not be as tidy as a single-platform campaign.

Two things follow. First, keep creative naming disciplined from day one — a variant ID that travels from your generation tool to the ad nameplate to your own analytics is the only way to attribute a conversion back to a hook. Second, decide in advance what "learning" means for the pilot: treat the first quarter as a measurement build, not a performance verdict.

The prep list before access opens

Pick three to five SKUs with proven margin and existing purchase intent. Conversational inventory rewards products people already search for.
Write the test matrix before the creative. Five hooks against two formats is ten ads; decide what each one is proving.
Produce the variants cheaply. Ten ads at agency rates is a quarter's budget; ten ads from a URL-to-ad pipeline is an afternoon.
Keep one offer per variant, and keep the offer legible without audio.
Check your rights. If you generate creative with an AI tool, confirm the plan includes full commercial rights before it runs as paid media.

What it costs to be ready

This is the part we can talk about with our own numbers. Prizmad turns a product URL into a finished UGC-style ad — script, hooks and offer copy written, one of 50+ avatars lip-synced to an AI voiceover, captions and music composited — in about five minutes. Token math is flat and published: ≈1,000 tokens per finished ad, on plans of $39/mo for 3,000 tokens (≈3 ads), $99/mo for 8,000 (≈8), and $249/mo for 35,000 (≈35), with full commercial rights included.

The ten-variant matrix above is roughly one month on Starter at about $12 per ad — budgeted before you click generate, and independent of whether the pilot has opened to you yet.

The honest caveats

This is a U.S. pilot with select advertisers, and buying starts with managed service — there is no published date for when self-serve access through Amazon DSP opens more widely.
Amazon and OpenAI are not the only route in: more than 50 ad-tech partners already offer ChatGPT Ads access, and self-serve is live in a long list of markets.
If you don't sell in the U.S., today's news changes nothing about your channel plan.
And no integration makes a weak offer work. Buy-side plumbing decides how easy it is to spend; the creative and the offer decide whether the spend returns.

The direction of travel is clear enough: conversational inventory is becoming a normal line item in normal ad platforms. The teams that treat it that way — with a creative pipeline that can produce variants on demand — will be running tests while everyone else is still waiting for an invite.

Start with the creative side: turn a product URL into your first ad in about five minutes, or check what AI video actually costs before you budget the matrix.

If you want the buy-side timeline in order, read the two earlier posts: the $1B run rate and self-serve launch and the European expansion.