The new ad format is the headline. The partner list is the story.
On October 5, OpenAI announced visual ads that appear during image generation in ChatGPT, with US testing starting later this month. In the same post it listed ten attribution partners, three advanced measurement partners, three incrementality partners, three data integrations and two brand safety firms.
That is not a product launch. That is a platform applying for a line in the media plan.
What actually shipped on October 5
The visual format shows product images next to the picture a user is generating. OpenAI says the ads are clearly labelled, stay separate from the generated image, and do not influence answers.
The measurement side is wider. Conversion data can now flow in through Hightouch, Tealium and LiveRamp. Click attribution and Conversions API support covers AppsFlyer, Triple Whale, Adjust, DV Rockerbox, Northbeam, Branch, Singular, Kochava, Airbridge and Tenjin. Fospha, Measured and INCRMNTAL handle full-funnel modelling, and Haus, Measured and WorkMagic will run geo-based incrementality experiments.
On brand safety, OpenAI added Negative Phrases for qualifying advertisers and started evaluation pilots with DoubleVerify and Integral Ad Science. The independent firms test the safeguards in a controlled environment without reading real conversations.
The scale argument has not changed. OpenAI says ChatGPT reaches 1.2 billion people a week, and I covered how fast that turned into money when ChatGPT ads hit a billion-dollar run rate in 200 days.
Read the case studies like a buyer, not a fan
OpenAI published three early results. According to DV Rockerbox, WeightWatchers saw an attributed cost per acquisition 15.3 percent lower than its blended paid search benchmark. WorkMagic reported statistically significant lift for wellness brand Dose, with 67 percent of incremental purchases from net-new customers. Triple Whale said 93 percent of Portland Leather's visitors from ChatGPT Ads were new.
Those numbers are encouraging. They are also exactly what every new ad platform publishes in its first year.
Early inventory is cheap because competition is thin. Early advertisers are hand-picked and well supported. And "new visitors" is the easiest metric to win on a platform where almost nobody has advertised before, because almost every visitor is new by definition.
The only figure in that list that I would put in front of a CFO is the incrementality result, because it tries to answer the question that matters: would those purchases have happened anyway? Even that is one brand, one partner and one test design. It is a reason to run your own test, not a reason to move budget.
The measurement stack is a signal of intent
Here is why the partner list matters more than the format. A brand selling through Google and Meta already has its numbers flowing through some combination of these vendors. Northbeam, Triple Whale and Rockerbox sit inside many e-commerce dashboards. AppsFlyer and Adjust run most app attribution.
By plugging into all of them at once, OpenAI removes the main operational excuse for not testing. You do not need a new reporting process. ChatGPT shows up as another row in the system your analyst already opens every Monday.
That cuts both ways. It makes testing easy. It also means ChatGPT Ads will be compared directly, row by row, with paid search and paid social, using the same attribution windows and the same blind spots. A channel that lives in the shortlist moment, before the buyer ever types a branded query, can look weak in last-click reporting even when it is doing real work.
I made the case months ago that the shortlist economy is eating paid search. If that is right, the attribution model you choose will decide whether ChatGPT Ads look like a bargain or a waste, regardless of what they actually do.
Where the visual format fits, and where it does not
Placing ads next to image generation is a careful choice. Someone generating a picture of a living room or a wedding table is already imagining a purchase.
It also limits the audience. Image generation is a fraction of ChatGPT usage, and the categories that fit are visual by nature: home, fashion, travel, food, beauty, fitness. If you sell payroll software, this format is not for you yet, and the text formats remain your only route.
For the brands that do fit, creative becomes the variable again. A visual ad sitting next to an image the user just made has to look like it belongs in that moment. Stock product shots on white backgrounds will lose to images that show the product in use, in a setting that looks like the one being generated.
How I would test it this quarter
Start with a hypothesis, not a budget. Pick one product line where the buyer typically explores options before searching for a brand. That is the moment ChatGPT owns.
Set the test up as incrementality from day one. Use a geo split or a holdout, even a rough one, and agree the success metric before launch: incremental new customers at an acceptable cost, not attributed ROAS. Platform-reported numbers in the first months will be flattering for the reasons above.
Feed it clean conversion data through whichever integration your stack already supports, and set your own Negative Phrases before the first impression. Brand safety in a conversational environment is new territory, and the independent evaluations are still pilots.
Then compare the result with your paid search cost for the same customer. If ChatGPT finds them earlier and cheaper, the budget will move on its own. If it does not, you will have the data to say no with confidence when the sales call arrives.
If you would rather have a team design and run that test, difrnt. (difrnt.ro) runs this exact kind of paid media work for mid-market and enterprise brands.
OpenAI just built the plumbing that lets you judge ChatGPT like any other channel. Make sure you judge it on the right numbers.