ChatGPT Ads Look Like a Performance Channel Before They Are One

If you already run Amazon campaigns, there is now a route to spending that budget inside ChatGPT without building anything new. Amazon opened a pilot on 10 September that lets selected US advertisers extend existing campaigns into the chat interface, with Delta Vacations among the first brands testing it (PPC Land). European buying runs through Adform, which OpenAI named its first European-headquartered platform partner on the same day, with Volkswagen and Vodafone testing (PPC Land).

Open the reporting and it will look like every other channel on your plan: impressions, clicks, CTR, average CPC, a conversions column. The numbers are real, and the ads may well be working. What no part of that reporting can tell you is whether the conversions in the column would have arrived without the ads, and a column that looks like performance reporting gets funded like performance reporting. That gap, rather than any shortage of data, is what will misallocate money in Q4 plans.

Seven months from launch to a full performance dashboard

ChatGPT only started serving ads in the US on 9 February, as a pilot with a minimum spend somewhere between 200,000 and 250,000 dollars. Self-serve buying, CPC bidding, a measurement pixel and a conversions API that posts sales straight from your own servers all arrived together on 5 May, and conversion-optimized campaigns followed on 5 June (PPC Land). In August OpenAI added view-through conversions, which credit a sale that follows an impression nobody clicked (PPC Land).

So a seven-month-old channel now presents a dashboard with the same columns as one you have run for a decade, and those columns quietly borrow the credibility of the older ones. The Amazon announcement, meanwhile, said nothing about how attribution will work across the ChatGPT surface, which is a strange thing to leave out of a launch aimed at performance buyers.

What the conversions column actually counts

The view-through metric shows why that familiarity misleads. Those conversions run on a fixed one-day window from an impression, and that window operates independently of whatever click window the account is set to. OpenAI also keeps them out of the Conversions metric, which stays a click-through total (PPC Land). So two conversion numbers run on two different clocks, and the column most people read contains only one of them.

OpenAI does not publish what you would need to interpret either. How modelled conversions get produced, how much uncertainty sits around them, which impressions qualify for view-through credit: none of it is documented. Auction insights and share of voice do not exist here, and nothing tells you which prompts are surfacing a competitor instead of you. One independent analysis calls the result "the most self-measurable and least externally observable major ad surface in performance marketing" (greghal.no). The platform can always show you a number, and you have nothing outside the platform to check it against.

A conversion in that column is a conversion that happened after an ad. Whether it happened because of the ad, which is what incrementality asks, is a different question, and none of the reporting is built to answer it. In an established channel you at least have years of your own holdout results telling you roughly how much reported volume tends to be genuinely caused. Conversion-optimized campaigns here have existed since June, so nobody has that history yet, OpenAI included.

Everyone selling inventory is also shipping the scorecard

It would be easier to wait this out if the gap looked like one young platform's growing pains. In the same week, Google shipped the clearest example yet of a seller grading its own work: Data Strength Uplift, a metric that sits in Google Ads next to your campaigns and calculates "the additional conversions recovered by an advertiser's first-party data setup" (PPC Land). Google decides what would have happened without your data setup, measures against that, and reports the difference as recovered conversions.

The supporting numbers have the same shape. Google's headline 14 percent conversion uplift for its tag gateway carries a footnote scoping it to Google's own data on the finance vertical, comparing the second half of 2024 against the first half of 2025, while the body text repeating that figure carries no vertical qualifier at all.

Where the money should come from

Buy the channel anyway. Just be precise about which budget line pays for it, and build the one measurement the platform cannot hand you.

Fund it from test and learning budget rather than the performance line, and treat the conversions column as a pacing signal, something that tells you delivery is happening and roughly where, not as a return you can defend in a QBR. Then run the comparison yourself: keep the campaign out of a set of regions, run it everywhere else, and measure the gap in total orders rather than in platform-reported conversions. That gap is the only number in this channel that does not originate with the company selling you the inventory. Keep US and European results apart while you do it, because European targeting draws on "the current conversation topic, general location and device type only", with past chats and stored memories excluded (PPC Land), so the two are not selecting on the same signals and should not be pooled into one read.

Moving slower than the discourse suggests is safer than it sounds. The IAB's September update, drawn from more than 200 brand and agency decision makers, puts optimizing content for AI-generated answers at the top of the list of areas getting more attention, at 76 percent, while only 16 percent of brands systematically track their visibility inside AI-generated answers at all (PPC Land). Stated priority is running a long way ahead of operational capability across the entire buy side.

Our own hiring data says the same thing more bluntly. The jobs pipeline behind Promptafire's skills tracker reads marketing postings from 152 companies' applicant tracking feeds: 4,610 marketing roles since April, 2,091 of them still live at the 13 September scrape. Across all of them, exactly one names ChatGPT Ads as a channel the person hired will run, and it first appeared on 5 September. Twenty-six companies have asked for answer engine or generative engine optimization, and every one of those requests sits on the organic side. The paid half of AI answers has not produced a single dedicated role yet.

The advice has a real hole in it. A regional holdout needs enough volume to detect an effect, and pilot budgets on an invite-only US surface usually will not have it. For most teams the first quarter in this channel buys a position and a learning curve, not a measured return. That is a defensible thing to spend money on, but it has to be asked for in those words, because the conversions column will not say it for you.

Sources

• Amazon Ads gains a ChatGPT ad surface as Delta Vacations tests first, PPC Land, 10 September 2026
• Adform gains ChatGPT Ads access for clients as OpenAI enters Europe, PPC Land, 10 September 2026
• OpenAI's ChatGPT ads are getting conversion optimization, PPC Land
• ChatGPT Ads gains view-through conversions with a fixed 1-day window, PPC Land
• ChatGPT Ads measurement: what OpenAI will not show you, 28 July 2026
• Google's Data Strength Uplift metric quantifies first-party data in Google Ads, PPC Land, 10 September 2026
• IAB lifts 2026 US ad forecast 2.8 points to 12.3 on strong first half, PPC Land, 10 September 2026
• IAB raises 2026 U.S. ad spend forecast, IAB, 10 September 2026