ChatGPT Is an Auction the Midterms Can't Enter. Test It With a Holdout.

On October 5, OpenAI said it would start testing a new ad format in ChatGPT later this month, with pictures from advertisers appearing while someone generates an image, labeled as ads and kept apart from the image the person asked for. The test runs in the US with a starting group of advertisers, and only people on the Free and Go plans will see it.

The same announcement brought the measurement a serious buyer needs. Advertisers can now send conversions through Hightouch, Tealium and LiveRamp, read results through ten attribution partners including AppsFlyer and Northbeam, and run geo-based lift tests with Haus, Measured and WorkMagic. OpenAI calls its own incrementality work early, which is the most honest line in the release.

The timing is why I would pay attention this month. In August I wrote that political money would come for your Q4, and it has. AdImpact's June update projects $11.6 billion in political advertising this cycle, the most expensive midterm on record, with $2.7 billion of it on connected TV. That money is not price-sensitive, and most of it lands in the final weeks, which means the next eighteen days. It is bidding against you on broadcast, connected TV, YouTube and Meta right now.

It cannot bid inside ChatGPT. In May, OpenAI said it "will not allow advertisers to run political ads on our platform this cycle," and its ad policy bars ads that advocate for or against political actors, elections or public policy. That puts ChatGPT among the few large ad surfaces where campaign money has no seat. Retailers still have one, and holiday demand will push prices up there too, but the least price-sensitive buyer of the season is locked out.

The price history is worth knowing before you build a plan. ChatGPT ads launched on February 9 at about $60 per thousand impressions, with minimums around $200,000. By April, buyers were reporting CPMs as low as $25. In May, OpenAI opened a self-serve Ads Manager, in beta, to US businesses, with cost-per-click bidding. New inventory is cheapest while advertisers are still deciding whether to show up. I was among the first ten Facebook Preferred Marketing Developer agencies in the world, and the advertisers who learned that auction while it was cheap kept what they learned after it wasn't.

Cheap inventory makes a test inexpensive. It does not tell you whether the channel works for you, and neither do the results in OpenAI's announcement. WeightWatchers' attributed cost per acquisition came in 15.3 percent below its blended paid search benchmark, and 93 percent of Portland Leather's visitors from ChatGPT ads were new to the site. Both numbers come from attribution, and a new channel looks efficient in attribution because nothing else in your stack is claiming its conversions yet. The Dose result, a lift test run by WorkMagic, is the kind worth copying. Run ChatGPT in some markets, hold others out, and compare total sales between them. You do not need a data science team to run that holdout.

Set up the plumbing before the first dollar goes out. Send conversions server-side through the Conversions API or a connector you already run, because a pixel alone misses purchases that happen later or on another device. When the visual unit opens up, build creative for how it works. It shows products in use and the experiences they make possible, so a white-background product shot is the wrong asset. Turn on Negative Phrases if your account qualifies, since the independent checks from DoubleVerify and Integral Ad Science are still in development.

Some of my clients cannot buy any of this. Alcohol is prohibited outright. Financial services, health and legal advertisers are reviewed case by case. The policy that keeps candidates out also bars ads that argue for or against public policy, so a trade association running an issue campaign gets nothing from Ads Manager. For those clients the way into ChatGPT is the answer itself, what the model says about the issue when nobody is paying for the space, and that gets earned through the sources it trusts. Your organization already has an AI reputation, whether or not you buy ads.

After November 3 the campaign money leaves every other auction you buy, and prices there will ease. The window inside ChatGPT closes on a different clock, set by every competitor who read the same announcement you did. I would rather walk into January with a clean holdout result than with someone else's case study.