Amazon Plugs Its Ad Budget Into ChatGPT: AI Answer Inventory Now Has a Reserve Price

1. The Event and Deeper Analysis

On September 10, 2026, Amazon Ads announced a partnership with OpenAI, letting select U.S. advertisers run ads in ChatGPT through Amazon DSP. Delta Vacations is among the first testers, with President Katrin Koenig named. The same week, ChatGPT's ad business was reported at a $1 billion annualized run rate, under 200 days after its February 9 launch.

Amazon's Q2 ad revenue was $19.8 billion, up 26% year over year, the largest retail media network in the world. Yet it had used robots.txt to block OpenAI's crawlers and Google's shopping agents. The logic is not hard to reconstruct: retail media grows on sellable inventory, ad density inside Amazon's own store is near its ceiling, and over the past year it has pulled Netflix, Spotify, Disney and ESPN inventory into its DSP. ChatGPT's 900 million weekly active users are attention Amazon cannot grow for itself. Blocking protects the data, not the growth curve.

Peer platforms are on the same road at different tempos. Google took AI Max out of beta on April 15 and retired Dynamic Search Ads, then on September 1 force-converted broad match campaigns into AI Max, narrowing the route to AI answer placements to a single door. Microsoft opened AI Max to every account globally on August 19, explicitly naming conversational queries in Bing and Copilot. Perplexity went the other way, dropping ads in February for subscriptions. On OpenAI's side Amazon is not the only pipe either: Criteo, StackAdapt and several agency holding companies were already connected. What makes Amazon unusual is that it has no assistant traffic of its own, yet it plugged the demand-side money in.

This differs from the product feed story of September 6: that was the supply-side ticket, this is the demand-side money pipe opening. One question remains — once inventory has a reserve price, which queries get bought, which do not, and which side Taiwanese SMBs should bet on.

2. Event Details and the Full Numbers

This is not a new self-serve placement type. ChatGPT inventory sits under Amazon DSP's existing omnichannel supply — the executive fronting the announcement, Chris Conetta, is Director of Omnichannel Supply at Amazon DSP. Advertisers extend campaigns they already run; Amazon reps operate them and use "context hints" to match ads to topics and keywords; OpenAI's system decides which creative appears; units land as text or images beneath the response. Audience signals come from Amazon's own browsing, shopping and streaming data, without ad identifiers.

Measurement is the obvious gap. Inside its own estate Amazon can name which retailers converted through Omnichannel Metrics and has stretched the Marketing Cloud lookback window from 13 months to 25. How that closed loop attaches to someone else's conversational interface, the announcement does not say.

Where does the reserve price come from? On one side, $19.8 billion of quarterly cash flow growing 26%. On the other, an entry cost falling fast: at the February launch the minimum commitment was $200,000 at a $60 CPM; by April the self-serve interface had cut the minimum to $50,000 and CPMs to as low as $25; by May CPC bidding arrived and the floor reached zero for some buyers.

Sensor Tower's September 1 report: in August, ads per user per hour on the U.S. ChatGPT mobile app rose 163% versus April, averaging 26% month over month, with roughly 1,200 unique U.S. advertisers active, while shopping and retail's share of U.S. ad spend halved in four months.

3. Biddable vs Non-Biddable: A Test for Sorting Queries

Split the query list in two before budgeting. Ask four questions of each:

• Does the category carry retail media budget? No bidders, no auction.
• Does it point at a transactable SKU? If it cannot collapse into an item, an ad has nothing to attach to.
• Does it need local or regulatory context? The more it turns on statutes and effective dates, the less a platform lets an ad answer.
• Is lifetime value high enough? Margin below the cost of one impression and nobody competes.

First, second and fourth yes with the third no: citations get diluted within six months. Invert that and the query cannot be bought.

4. Immediate Actions for Three Audiences

(a) Brand owners

• Sort the product list with the four dimensions inside a week: biddable gets money, non-biddable gets content.
• If you sell electronics, appliances, travel, beauty or consumer finance, move "AI answer exposure" into the media budget this quarter and test it with 3% to 5% of digital spend.
• With a U.S. entity or distributor, ask your Amazon rep about pilot access and rates; without one, stop waiting for a Taiwan launch this year.

(b) Marketing operators

• Build a 30 to 50 prompt list and run it weekly across ChatGPT, Gemini and Copilot — about 90 minutes.
• Record whether the mention sits above or below the ad unit, not just how often; position moves two to three months earlier than count.
• Check category rotation monthly. The month yours enters the ranking, pull that quarter's content schedule forward.

(c) Developers

• Audit robots.txt and llms.txt so OAI-SearchBot and Google-Extended are deliberate choices rather than defaults — half a day.
• Complete Product, FAQPage, Service and LocalBusiness markup and bucket referrers so AI sources are not folded into direct — two to three days.
• Turn the prompt list into a scheduled script against the official APIs writing to a database: a few dollars a month, with replayable history.

5. AI Visibility Monitoring Tools Compared

Most GEO tools track organic citations only and cannot show ad-unit squeeze:

6. What Nobody Will Tell You

First, the denominator behind 163% is a problem. It measures ads per user per hour, and the denominator is time spent, while time spent and session structure are themselves changing — longer conversations and more turns fit more trigger points into the same hour. The baseline matters too: that report's mobile data starts April 1 and desktop June 1, so "up versus April" compares against the first month mobile was measured at all.

Second, the category rotation is a change in advertiser mix, not simply more load. Shopping and retail fell from 37% to 21%, financial services went 2% to 13%, travel 5% to 11%. Volume rose while shopping's share halved, meaning incoming finance and travel buyers enlarged the pool. A rising share in your category reads as competitors arriving, not opportunity expanding.

7. The No-SaaS-Subscription Alternative

A version you can run yourself on NT$0 to NT$500 a month:

• Fix 20 to 30 prompts, half biddable and half not, and do not change them for six months — change them and you lose the baseline.
• Run them in batch against the official APIs, once per engine per prompt, one round a week: under US$1 a round at current text model pricing, roughly US$3 to US$5 a month.
• Log to a spreadsheet: date, platform, ad unit present, whether you were mentioned, above or below the unit. About 30 minutes of checking.
• Compare the two groups' citation rates monthly; divergence means dilution has begun, and content hours all move to the non-biddable group.
• To automate, run a fetch script on a Laravel schedule: one to two days to build, under an hour a month to maintain.

8. Frequently Asked Questions

Can Taiwanese advertisers buy ChatGPT Ads?

No. The pilot is limited to select U.S. advertisers, and OpenAI's own ads are not open here either.

Wasn't Amazon blocking ChatGPT?

What it blocks is crawler access to its store pages; what it opened is the ad-buying interface.

How do I measure how much I have been diluted?

Track position, not count — whether you land above or below the ad unit. It moves two to three months earlier.

Will AI ads swallow the whole search budget?

Not soon. eMarketer puts U.S. standalone chatbot advertising below $1 billion in 2026 and only just above $5 billion by 2030, and more than 80% of AI advertising sits next to AI content rather than inside conversations.

9. My View

The market reads this as "ChatGPT is about to be stuffed with ads." That misses the point. The turn is that inventory now has a reserve price: introduce an auction and you get a clearing price, and anything with a clearing price eventually goes to the highest bidder.

But auctions only happen where bid cost sits below expected return: retail media can price a travel package's conversion value, and cannot price "which firm in Taichung handles cross-border e-commerce tax filings well." Most ScriptWalker clients sit in the latter.

10. Sources

Primary

• Amazon Ads announcement, Sensor Tower, Profound pricing

Third-party

• MediaPost, PPC Land (Amazon surface), PPC Land (CPM compression), PPC Land (Microsoft AI Max), PPC Land (eMarketer forecast)
• Digiday, Marketing Dive, Modern Retail, eMarketer, Search Engine Land, Variety