If you sell on Amazon, here is the sentence to take away from this week: Amazon will now help an advertiser put an ad in front of ChatGPT's users, and it still restricts OpenAI's software from reading your detail page. The paid way into the conversation runs through Amazon. The organic way in, where a model reads your listing and decides to mention you, is a door Amazon has spent the last year pushing shut.
That's not a scandal. It's a business model, and it's worth understanding before somebody sells you a retainer built on it. Nothing about your September P&L moves because of this. What changes is who controls how you get discovered in chat, and the answer this week became "Amazon, both ways."
What happened
On September 10, 2026, Amazon Ads announced a partnership with OpenAI letting select U.S. advertisers extend Amazon DSP campaigns into ChatGPT as a pilot called ChatGPT Ads. Delta Vacations is one of the first testers. Per Marketing Dive, it is sold as a DSP managed service with CPC or CPM pricing: text and image units under ChatGPT responses, product feed ads that build creative automatically from a catalog, aggregated reporting, and OpenAI keeping "all ad delivery decisions."
That's the news. The rest is what it means if your business is a catalog on Amazon.
Why most brand owners will read this wrong
The dumb take: "Amazon and OpenAI are partners now, so ChatGPT is about to become an Amazon shopping channel and I need to be in it."
No. Read who launched it. The first named advertiser is a vacation company, the buying path is managed DSP, and the pilot is invite-only. This is Amazon selling reach to big advertisers, including ones that don't sell on Amazon at all. It is not a Sponsored Products extension, and no public material says an ad can point at a third-party seller's detail page. If your whole ad account is SP and SB, this product doesn't exist for you yet.
The mirror take: "It's a DSP thing, ignore it."
Also wrong, and the more expensive mistake. Two facts sit next to each other this week and nobody in ecommerce coverage is putting them together:
- Coverage of the deal notes that Amazon had been restricting ChatGPT, Claude and Gemini from its store while it built its own shopping tools, including blocking OpenAI's live-retrieval agents in robots.txt.
- According to Sensor Tower data cited in the coverage, Amazon was the top retail advertiser on ChatGPT between April and August, and ad density on the ChatGPT mobile app rose 163% over the same period.
Put them together: Amazon keeps the model from reading the shelf, buys its own placement in the answer, and now resells that placement to other advertisers. The real signal is that chat discovery for an Amazon catalog is turning into a paid surface brokered by the company that owns your listing. I've written twice about chat discovery (in June on ChatGPT ads and in August on the site: operator). Both times I assumed the organic route would matter most. This week makes the paid route the one with a salesperson attached.
What actually changes at $200K/mo
1. You're not in the pilot, and you probably won't be this year. Managed DSP is a relationship, not a checkbox. Most brands doing $200K a month on Amazon run SP, SB and maybe some self-serve Sponsored Display. When this reaches you, it'll arrive through DSP or through an agency with DSP access, and the minimum spend will decide whether you're in before any strategy does. Plan on that, not on a button showing up in your console.
2. Product feed ads mean your catalog data is the ad. This is the part that will reach catalog brands. A feed ad builds creative automatically from product data, and for an Amazon brand that means your title, your main image and your price. Nobody briefs it, and nobody reviews each version. We already live with this: the 75-character title rewrites go live on a 14-day review clock if nobody on your team is watching. A feed ad in a chat answer is the same default showing up somewhere new. If your first 75 characters say "Premium Quality Ultimate" and your main image sits at 60% frame fill, that's what the machine will show a shopper who asked a specific question.
3. The answer page is filling up with ads. Up 163% in four months, per Sensor Tower as reported. Every paid unit under a response takes space from the organic mention. A brand with no ad budget that a model happens to name is now competing below an ad in the same scroll. Unpaid mentions still count. They just take up less of the answer page every quarter.
4. You get reach reporting, not sales reporting. The disclosed reporting is aggregated impressions, clicks, cost per result, CPM and CPC. PPC Land points out that nothing public explains how a chat impression gets deduplicated against other DSP placements or tied to a purchase. For a brand that grades ads on cost per unit sold, that makes it an awareness buy. It's also being sold by a company whose Sponsored Products auction pricing is the subject of an FTC complaint filed August 31. That complaint is allegations, not findings. But the week somebody asks you to trust a new surface's pricing isn't the week to skip reading its terms.
5. Your only readout is still branded search. An Amazon seller has one place to see chat exposure: branded query volume on Amazon that rises with no ad spend or social push behind it. That was true in June. The difference now is that a competitor on DSP can start driving that lift with money. So an unexplained branded lift on a competitor's name in your category deserves a note in the log with a date.
What I'd do this week
- Ask your Amazon Ads rep or agency one written question. "Is ChatGPT Ads available to us, through which buying path, can a unit link to a detail page, and what purchase attribution exists?" An answer with a buying path and a yes or no on attribution means somebody checked. A paragraph about the future of conversational commerce means nobody did, and that reply is the answer.
- Don't move a dollar of SP or SB budget. If you already run DSP with an off-Amazon destination and your rep offers the pilot, write the stop rule before you start: a spend cap, a date, and the branded-search lift that would justify a second month. No stop rule, no test.
- Read your top 10 listings the way a feed reads them. Export title, main image URL and price to a sheet and look at them together, with no bullets, A+ or reviews. Nine weeks before peak, that row is your ad in any automated placement, on Amazon or off it. Fix what reads as filler in the first 75 characters and re-crop any hero under 80% frame fill. It's an afternoon on files you already own, and it's correct whether or not this pilot ever reaches you.
- Mark September 10 on your branded search dashboard. Pull Search Query Performance for your brand terms and your top two competitors' brand terms, save it with the date in the filename, and diary a comparison for the first week of November. If a competitor's branded volume jumps and nothing on Amazon explains it, you'll know where to look.
- Don't start anything new before January. Nine weeks from peak is the wrong time to add a new paid surface you can't measure. Run the question in Q1, with a quarter of pilot results other people have paid for.
What I'd ignore
"Amazon is opening up to AI agents." An ad deal isn't a crawler policy. Nothing announced changes whether OpenAI's software can read your listing. If anything, it shows Amazon prefers getting paid for visibility in chat over allowing it for free.
The $1 billion run rate. That's OpenAI's revenue milestone, not a performance benchmark for your category.
The "fastest-growing engagement opportunity" quote. It came from the person whose job is selling the inventory. File it with every vendor quote about their own product.
Anyone pitching "ChatGPT Ads management" to an SP-only brand this fall. There's nothing for them to manage. The pilot is invite-only, managed by Amazon, and delivery decisions belong to OpenAI. A retainer for that is a retainer for waiting.
The urge to build something "for AI." The work that makes your product show up correctly in a feed ad is the same work that makes it legible to Amazon's own assistant: a clean title, a hero that works at thumbnail size, complete attributes, and A+ copy that makes its argument in text. I'd have told you to do all of that anyway.
Last month the question was whether a model could read your listing. This month Amazon added a second question: whether you'll pay to appear next to the answer. It owns the answer to both.