1. The potential reach is huge, even without consent for personalized targeting
OpenAI switched on advertising in ChatGPT in France this month, and in 30 other countries. But not for everyone. Like Netflix before it, the company has kept ads to users of the free tier and the Go plan. Plus and Pro subscribers, along with Enterprise customers, are still spared.
ChatGPT says it has close to 900 million weekly active users and only 35 million paying subscribers. That gap gives a sense of the pool OpenAI could theoretically monetize, even if not all of those users are reachable by advertisers right away.
European rules pushed OpenAI to use consent as its legal basis for personalized advertising. Users who opt in can be targeted on their ad history, their interests and data advertisers pass back to the platform.
Users who refuse still see ads. But the selection then rests on the conversation itself and an approximate location, without touching memory or history. OpenAI can keep monetizing those users through contextual targeting. The company says 20% of interactions carry commercial intent, whether research, comparison or a planned purchase.
The product has changed a lot since the U.S. pilot launched in February. Back then ChatGPT Ads sold at around a $60 CPM, with a $200,000 minimum commitment. That floor dropped to $50,000 in April, before the launch of an Ads Manager with no spending minimum.
Advertisers can upload their ads there, set a budget and manage their bids. In France, access to the tool is still limited to agencies owned by the big holdcos.
Independent advertisers and agencies have to go through one of OpenAI's five adtech partners: Adobe, Criteo, Kargo, Pacvue and StackAdapt. They handle creative, budgets and bidding, while OpenAI keeps control of delivery.
OpenAI has also added CPC buying, with bids initially seen between $3 and $5. The platform now offers conversion optimization: the algorithm no longer chases an impression or a click, but an action the advertiser defines. Geotargeting and custom audiences round out the setup.
Targeting is still harder to control than on established platforms. Advertisers currently supply "context hints," a description of the conversations, topics or keywords where their product may be relevant. Those hints guide the model, but they do not work as exact targeting criteria.
OpenAI is now testing negative targeting instructions with a small group of advertisers. They would let advertisers specify the conversational contexts they do not want to appear next to. The distinction matters: OpenAI describes this as "guidance" given to the system, not a guaranteed exclusion comparable to negative keywords in Google Ads. The company has shared no rollout timeline.
The move answers several criticisms reported by Adweek. Advertisers struggle to describe the audience they want, have few ways to control where their ads run and get little visibility into the actual environment around their impressions. In six months, OpenAI has assembled the main building blocks expected of an ad platform. Having them available does not yet mean they deliver the level of control, transparency or maturity of Google and Meta.
3. Product feeds now matter more than individual creatives
The original format paired a headline, a description, an image and a link. It has since gained a larger image and customizable buttons, then e-commerce formats showing prices, reviews and several products in a carousel.
With these dynamic formats, the advertiser no longer builds every ad. It hands over its catalog, and OpenAI picks the products it judges relevant to the conversation, much like Google Shopping. The model makes two calls that media buyers largely made until now: whether to show an ad at all, and which product to push.
The unit of optimization shifts from the ad to the feed. Feed quality now matters as much as bids or creative: clear titles, categories, attributes, images, prices and availability. A product risks being dropped, or served in the wrong context, if it is poorly described, filed in the wrong category or wrongly flagged as in stock. Teams will have to bring together skills that have sat apart in search, retail media, e-commerce and dynamic creative.
Handing that autonomy to OpenAI also raises a control problem. To understand performance, media buyers need to connect the conversation context, the product served and the conversion that followed. Without that level of reporting, they will know a campaign works, but far less about which signals led the model to favor one product over another.
4. OpenAI is already preparing to open up to SMBs
In late July, Digiday spotted an offer giving $50 in credits after a matching spend within 14 days of opening an account, along with a $100 version. Beyond the amount, the mechanism marks a shift from launch mode to acquisition mode. OpenAI now wants to turn signups into first budgets, then those tests into recurring spend.
The natural target is small and midsize businesses. Large advertisers can bring big budgets fast, but there are only so many of them. Google and Meta built their ad businesses on a long tail of millions of companies able to run campaigns on their own with a few hundred or a few thousand dollars. Neither breaks out revenue by advertiser size, but SMBs are generally estimated to account for about two-thirds of their ad revenue. Of the $491 billion Google and Meta generated in 2025, that would be roughly $330 billion.
The stakes are higher still because OpenAI is reportedly targeting $100 billion in ad revenue by 2030. On a Google or Meta-like split, close to $65 billion would have to come from SMBs. More of them would also thicken auctions across more categories, more geographies and more conversational contexts. Without that depth of demand, ChatGPT Ads would stay dependent on test budgets from a few hundred large brands.
So OpenAI is building an SMB team, hiring in data science, growth, demand generation and sales operations. Part of the selling could go to outsourced vendors, as it does at Google and Meta. The goal is to run the whole funnel at scale: recruitment, first campaign, spend ramp and retention.
The promo credit is only step one. To convert that long tail, OpenAI will have to let a company with no media team build a campaign in minutes, understand the results and optimize toward a business goal. The credibility of its $100 billion projection will rest largely on turning millions of small tests into recurring budgets.
5. Measurement works now, but the proof of performance is still missing
OpenAI's measurement pixel is a snippet placed on the advertiser's site. It ties a click to a visit, a lead, an order, a subscription or a trial. The Conversions API sends the same events straight from the advertiser's servers, with less loss to cookies, browsers or ad blockers. Those signals also feed conversion optimization, so a badly configured setup degrades both the measurement and the bidding.
The platform offers post-click attribution and, for some accounts, a one-day post-view window. Post-view stays walled off in reporting: CPA, billing and optimization still run on post-click conversions.
OpenAI is also opening measurement to third parties. Fospha, Hightouch, LiveRamp, Triple Whale and WorkMagic are among the documented integrations for passing back web, app or offline conversions. Kochava says it is working with OpenAI on a cross-channel read. Verification firms such as IAS or DoubleVerify could come in on different ground, covering viewability, human traffic and brand safety, without measuring business impact.
The real gap is causality. The pixel and the Conversions API show that a conversion happened after an exposure or a click, not that the ad caused it. That risk runs high in ChatGPT, where the ad often appears in the middle of a journey already loaded with intent. The platform can then claim a sale that would have happened anyway.
The reverse is true too. This kind of measurement barely captures an ad that is remembered, then followed by a branded search, a direct visit or an in-store purchase. Buyers will have to cross OpenAI's reporting with their analytics, their CRM data and their offline sales, then run incrementality tests. Marketing mix modeling can fill in once the volumes and the history are there.
Marketplace Pulse founder Juozas Kaziukėnas' own six-month review confirms the limit: he found no robust ROI shared publicly. One of the few available cases shows decent click-through rates but very weak conversions, or none at all. ChatGPT Ads can now be tracked, attributed and optimized. Whether it scales depends on proving that the sales it claims would not have happened without the ads.
