100,000 advertisers tried ChatGPT Ads. Can they prove ROI?

By Nathan Brooks

More than 100,000 advertisers have tested ChatGPT Ads, but early evidence shows a wide revenue gap: $0.89 per U.S. session, against $2.45 for Google Search Ads.

That gap is the central issue for OpenAI as it tries to move the product from experimentation into permanent media plans. OpenAI must prove returns, not just distribution.

OpenAI launched ChatGPT Ads about nine months ago, added formats, partnerships and tools, and reached a $1 billion annualized revenue run rate as of this summer. That rate projects the summer’s revenue pace across a year; it does not establish advertiser returns.

Early conversations with CMOs and data specialists produced a mixed verdict. Twilio CMO Chris Koehler described the company’s participation as dabbling, while Intuit CMO Thomas Ranese called its work experimental. Neither signals a durable budget commitment.

How strong is the early evidence?

A working paper under review by researchers Maximilian Kaiser and Christian Schulze compared the systems across 102 advertiser sites. ChatGPT generated $0.89 per U.S. session, while Google Search Ads generated $2.45. Revenue per session captures money generated for each visit, not profit.

“What makes us cautious about ‘it will come with time’ is the size of the gap,” Kaiser and Schulze said. Their warning targets the scale of underperformance, not merely the product’s nine-month age.

Adthena, a search intelligence platform, tested 30 enterprise brands from July through September. Fifty-two percent struggled to recoup their ChatGPT ad spend; 26% matched or beat Google ads; the rest reported mixed results. In other words, most did not recover the advertising cost.

OpenAI supplied a stronger counterexample in an October case study about Evernote. The note-taking app’s ChatGPT Ads produced a trial-to-paid conversion rate about 10% to 20% above its historical average across paid-acquisition channels earlier that year.

Trial-to-paid conversion measures the share of trial users who become paying customers. Because OpenAI published the case study, it remains a vendor-published result rather than the broader comparisons above.

Early evidence supports controlled tests, not a wholesale move from search budgets.

Can ChatGPT deliver enough relevant ad inventory?

Three constraints cap OpenAI’s ad growth: the size of ChatGPT’s free user base, the shortage of ad-relevant conversations, and the number of ads users will tolerate before leaving. More inventory lifts delivery, but excessive ads could damage the experience that made ChatGPT useful.

OpenAI’s head of global ad solutions, Dave Dugan, said in a June interview that ad delivery and volume improved as the company’s confidence in the ChatGPT Ads user experience grew.

Visibility remains uneven. Qualcomm CMO Don McGuire said large language models struggle to find audiences unless a brand’s targeting is extremely broad. Narrow categories therefore struggle to appear at scale, weakening return on investment, or ROI.

Kaiser and Schulze suggest moving ads from cards beside conversations into the answers themselves. That could improve context, but OpenAI’s ad principles say ads will not influence answers and that user experience outranks revenue. The trade-off is clear: relevance versus a firm product boundary.

OpenAI can expand inventory, but narrow targeting and its answer-integrity rules limit how far the format can travel.

Can better measurement prove the value?

Adthena CMO Ashley Fletcher calls measurement a recurring pain point. OpenAI has added a measurement pixel and conversion bidding. Advertisers can target a cost per acquisition, their target price for an acquisition, or a cost per click, the price of a visit.

Fletcher expects the tools to give CMOs cleaner numbers for deciding where to allocate budget, but fears they may expose a slower pattern.

ChatGPT ads may mostly nudge people at the first stage of brand consideration rather than produce the quick conversions many advertisers expect from other platforms and from ChatGPT. Fletcher points to Black Friday as a major test.

Cleaner attribution will decide whether the channel deserves a larger budget, not just a test line.

What is working now?

AdRoll vice president of product Hans Fischmann tags every ChatGPT ad the company runs and feeds the data into its broader media buy. His example: someone asks ChatGPT about a hotel on Tuesday, then sees the brand on television Thursday. He calls this the full-funnel play; in his view, the sequence connects research with later exposure and earns ChatGPT a permanent plan slot.

Natural Intelligence operates comparison sites Top10.com and BestMoney.com alongside a performance-marketing business. Head of strategy Jonathan Vardi moved ChatGPT Ads from a test budget into a growth budget, a repeatable category where spending rises as performance improves.

Vardi’s team lists the pain points each product solves, then turns them into specific ads that can appear while users ask ChatGPT difficult questions. Accurate post-click conversion data is more important still.

“If you don’t tell ChatGPT what users are doing on your site after they click your ad, you’re not allowing the ads algorithm to learn which users actually engage with what you have to offer and become paying customers,” Vardi said. Measuring only views or clicks, he added, constrains ROI as competition on the channel increases.

The workable pattern is clear: treat ChatGPT as a measured funnel, not a reach channel. Teams should expect the next planning cycle to hinge on whether they can trace ChatGPT exposure through to revenue, not on reach or click volume.