ChatGPT Ads Has Measurement. What It Doesn't Have Is a Referee.

There's a line near the end of OpenAI's announcement this week that most coverage treated as boilerplate. Having just taken ChatGPT Ads into Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan, the company said it would spend the next few months continuing to build "new formats, optimization tools, and measurement solutions."

It's tempting to read that as an admission that the platform can't measure anything yet. That reading is wrong, and it points the industry at the wrong problem. ChatGPT Ads has had a conversion pixel since an April pilot, a server-side Conversions API, conversion-optimised bidding since June, and one-day view-through attribution. A buyer in Singapore can run a campaign, count conversions and report a ROAS figure today.

The issue is who produces that figure. Every number in the stack comes from OpenAI, and the most-quoted proof point in the category comes from OpenAI's first ad tech partner. Nobody independent has checked any of it. That's a narrower problem than "measurement doesn't exist," and a more serious one.

The platform grades its own homework

What's missing is specific. There's no IAS, DoubleVerify or MRC accreditation on this inventory, so brand safety rests on OpenAI's own policy enforcement. There's no reach or frequency reporting and no age or gender breakdown. Device and country are the only cuts. There's no search-terms report either, and there won't be one, because advertisers are deliberately kept away from the conversations that triggered their ads. That's a defensible privacy choice. It also means a buyer can never see the context an impression ran against.

The consequences surfaced in MediaPost's reporting in late August. Nicholas Verity, who runs B2B lead-gen agency Cleverly, said OpenAI reported 57 clicks on his test while Google Analytics recorded fewer than 20 visits from the same campaign. He paused it. Peter Jaffray at Canadian agency Choice OMG ran three small campaigns, got more than twice the impressions he'd planned for, a click-through rate around 0.6% against the 2% he'd expect from search, and no buyers. According to TechTimes, OpenAI's own advertiser materials state the platform doesn't yet have performance benchmarks across advertisers, industries or campaign types.

None of that proves the channel doesn't work. A gap between 57 clicks and 20 sessions can come from tagging, consent loss or bot filtering on either side. What it shows is more basic. When the platform's number and the advertiser's number disagree, there's no third party to settle it.

The best number in the category isn't about the ads

The figure anchoring most of the bullish coverage, including an Inc. column this week, is that ChatGPT traffic converts about 1.5 times better than other channels. Its origin is worth tracing.

It comes from Criteo, which announced in March that it was the first ad tech partner integrating with OpenAI's ad pilot. The 1.5x was observed across a sample of 500 US Criteo retailers in February 2026, and it describes users referred from LLM platforms like ChatGPT compared with other referral channels. That's LLM referral traffic, much of it presumably organic, measured in a month when the ads pilot was days old and Criteo's integration hadn't rolled out. It was published by a company with a commercial interest in advertisers believing it.

Criteo didn't misrepresent anything. The footnote is right there. But a number about AI referrals, from one month, one market and one partner's client base, has been recycled into a claim about ChatGPT ad performance. That's how weak evidence becomes consensus.

The other headline stat has the same problem in a different shape. Comscore found sponsored ads in 47% of the ChatGPT conversations about video games it tracked in June, up from 3% in March. That's US desktop only, and it measures ad load, how often an ad appears, rather than whether anyone acted on one. Rising ad density is evidence OpenAI is monetising faster. It says nothing about whether the ads work.

Five holding companies, no proof required

dentsu, Havas Media, Omnicom Media, Publicis Groupe and WPP are all named as agency partners in the same release that says measurement is still being built. The likely explanation is that partner status is worth having whether or not any individual client campaign proves out: early access to formats, a line into OpenAI's product team, and a credential to take into pitches. This series has already looked at how much holding company economics still turns on privileged access to supply rather than proven client outcomes. This looks like another instance of it, though none of the partnership terms are public, so it remains a reading of incentives rather than a documented fact.

What's different in this region is that the holding companies aren't gatekeepers. Self-serve Ads Manager is live in all seven new markets from day one, and OpenAI lists technology partners as a third route in. For the long tail of Southeast Asian SMEs, the agency route is optional. That leaves an open question over what partner status actually changes for a client's campaign: better data access, earlier formats, or a logo on the credentials slide.

What the global coverage misses about Southeast Asia

The first gap is the audience. Ads only run on Free and Go plans. Plus, Pro, Business, Enterprise and Edu users never see them. OpenAI cites more than a billion weekly active users globally, but it hasn't published how that splits by plan in any Southeast Asian market, and the split is the only number that describes what's actually being bought. Total ChatGPT reach in a media plan is the wrong denominator.

The second is that Shopee got there first. Sea's marketplace has been available as an app inside ChatGPT across all seven launch markets since June, and Shopee is quoted as a partner in OpenAI's launch release. For brands already selling through Shopee, their products may be surfacing in commerce conversations without any media spend. What a paid placement adds on top of that is an incrementality question, and nobody has published an answer for this region.

The one genuinely new idea in all of this is Sponsored Agents, where a click opens a conversation with a business-sponsored agent inside ChatGPT instead of sending the user to a landing page. It's in limited testing with US advertisers. If it scales, the post-click journey becomes a conversation OpenAI hosts and reports on, which would widen the independent measurement gap rather than close it.

The burden of proof has shifted to the buyer

The practical effect of a channel with no referee is that verification becomes the advertiser's job. That changes what separates a serious test from an expensive one. Untagged ad clicks land in GA4 as chatgpt.com referrals, indistinguishable from organic mentions, so a brand that skips the tagging ends up inflating its organic AI numbers at the same time. Platform clicks that never reconcile against site sessions aren't a performance problem yet. They become one when budget decisions get made on them.

Southeast Asia has one structural advantage here. Seven markets launched at once, which is close to a natural geo experiment: some markets live, others held out, total sales compared rather than platform-attributed ones. It's the only incrementality read available that doesn't depend on OpenAI's numbers, and so far there's no sign that anyone in the region is publishing one. Until then, the case for moving established budget into this channel rests on evidence that can't be independently checked. That makes it test money, not reallocated money.

The timing is hard to ignore. The two companies built to referee digital advertising have spent this year changing hands, one to private equity and the other to Nielsen. The newest serious ad platform in Asia has launched with no referee at all. OpenAI says measurement will improve over the next few months, and it probably will. Until someone other than OpenAI can check it, that's a better scoreboard kept by the same team.