ChatGPT ads entered testing in February 2026 on the US free tiers 1, and by the last week of June OpenAI was pitching them from its first official Cannes Lions appearance 2. Experiment to sales pitch in five months. The format rewards the same assets that win everywhere else now, a clean product feed and copy that reads like an answer, so the testing decision is not complicated: ecommerce brands with strong feeds should run a small test, and most B2B and service accounts should wait.
What exists today, concretely
The practical question, how do ChatGPT ads work for advertisers right now, has a short answer: narrowly, in beta, and only in front of some users. Ads run on the US Free and Go tiers and have expanded to seven markets, most recently Japan and South Korea, while the Pro, Business, and Enterprise tiers remain ad-free 1.
The tier structure is the detail most coverage skips, and it should shape who bothers. The heaviest professional users, the ones on paid workspace plans, never see an ad. If your buyer is a VP evaluating software from an Enterprise seat, this surface does not reach them at any budget. If your buyer is a consumer asking a free-tier assistant which running shoe to get, it does.
Seven markets and two free tiers is not a channel yet. It is an inventory experiment with a sales team attached, and it should be treated with exactly that level of commitment.
The Cannes pitch, and how to read a seller’s numbers
OpenAI’s pitch numbers are strong, and every one of them comes from someone selling the inventory. David Dugan, OpenAI’s head of global ads solutions, called ChatGPT users “super intentional” and said roughly 20% of ChatGPT queries carry direct commercial intent 2. Criteo, an early partner, says more than 2,000 brands now run ChatGPT ads through its platform 2. And OpenAI says ad dismissal rates, the cross-out, are down 50% since launch 2.
Take the direction, discount the magnitude. Twenty percent commercial intent across ChatGPT’s query volume is a large pool of buying questions even if the true number lands lower once someone independent measures it. The dismissal figure deserves the most skepticism: a falling cross-out rate measures tolerance, not performance. Users clicking the X half as often tells you the ads got less annoying or better placed. It tells you nothing about whether they convert.
The Criteo number is the most useful of the three, because 2,000 brands buying through one partner means real budgets are moving and early benchmarks will exist soon. You do not have to be first. You do have to know when the data starts arriving.
Perplexity walked the other direction
While OpenAI built an ads business, Perplexity dismantled one. It phased out all advertising in February 2026, citing trust concerns, and now positions itself as the ad-free subscription alternative 3.
Two assistants, two opposite monetization bets, made in the same month. That tells you the category has no settled norms yet: nobody knows how much advertising an answer engine can carry before users stop trusting the answers. So do not build a plan that assumes this surface matures the way search did, on a predictable path from beta to auction to line item. The inventory, the formats, and possibly the entire ads program can change shape between quarters. Plans here should be written in pencil.
The feed is the creative
OpenAI launched product feed ads in its Ads Manager beta on June 11, 2026 4, and that launch tells you what the format wants: structured product data the model can read, rank, and present as part of an answer. There is no headline to write. The feed is the creative.
The same logic governs agentic checkout under the Universal Commerce Protocol, and it was the lesson of pMax before that. When the platform assembles the ad, the quality of your inputs is the only creative decision left. Titles that describe the product the way a buyer asks for it, complete attributes, accurate pricing, live availability: that is the campaign. A brand with a neglected feed buying this inventory is paying to have its worst asset read aloud by an assistant.
This is why the work we scope in the firm’s services keeps converging on feeds, offers, and landing pages regardless of which platform ships the new format. The targeting layer is automated everywhere now. What remains is what the machine reads.
Who should test, who should wait, and at what size
Test if you are an ecommerce brand with a clean, structured feed in a category where the buying question is recommendation-shaped: the “best X for Y” questions an assistant answers with specific products. Those are the queries where an ad that reads like an answer earns its placement, and where a strong feed does the selling.
Wait if you are B2B enterprise, because your buyers sit on the ad-free tiers. Wait if you are local services, because the recommendation-shaped demand for your category still runs through Maps and Search. And wait if your feed hygiene is not there yet, because you would be testing the platform with a broken input and learning nothing about either.
Measurement will be messy early. Assistant referrals blur across sessions and devices, early betas rarely ship clean conversion reporting, and the first months of any new inventory overweight novelty clicks. So size the test small enough to kill without a meeting: $2,000 to $5,000 over 30 days, one category of products, judged on whether it produces orders you can trace, not on whether the impressions feel exciting. The bar for when any new channel earns real budget in an account we run is the same one we describe in the FAQ: existing channels saturated first, incrementality over dashboard attribution always.
The last time an ad surface promised intent this clean, it was search, and the operators who won were the ones treating feed and data quality as strategy while everyone else wrote headlines. Whether ChatGPT ads become a durable channel or a Cannes talking point, the feed work pays either way. That makes the real decision this quarter not whether to buy the ads, but whether the feed is ready to be read.