Tools

ChatGPT’s new ads expand measurement, but the gains remain unproven

By Saint Zero Day · 2026-10-06 · 4 min read

OpenAI’s visual ads and measurement tools offer potential benefits, while evidence leaves questions about performance, reporting and privacy unresolved.

OpenAI announced on October 5 that it will test visual advertisements during ChatGPT image generation later this month in the United States, alongside expanded advertising measurement partnerships. The changes offer businesses more ways to assess campaigns. They do not yet demonstrate better shopping experiences or a general performance advantage over existing advertising channels. The visual format remains a planned test, and earlier campaigns have generated unresolved measurement complaints. [1] [3]

The proposed advertisements will depict product inspiration, usage or experiences, with labels separating them from the image being generated. OpenAI says ads do not influence answers and argues that advertising supports access beyond what subscriptions can fund. Editorial inference: that business model could benefit people who cannot afford subscriptions, while visual examples could help shoppers understand unfamiliar products. The announcement supplies no measured access or user-experience improvement from this expansion. [1]

Measurement connects advertising with subsequent actions. OpenAI’s website pixel and Conversions API accept purchase, lead, installation and sign-up signals, while integrations connect existing advertiser systems. These tools could help businesses allocate spending more effectively. OpenAI is also exploring geographic experiments intended to measure additional business caused by ads. That differs from attribution, which assigns credit to an advertisement preceding a purchase without necessarily establishing that it caused one. [2]

Why it matters

OpenAI reports that DV Rockerbox measured WeightWatchers’ attributed acquisition cost at 15.3% below its blended paid-search benchmark. Separately, WorkMagic estimated 2.3 times as many incremental Dose orders as last-click attribution captured. These comparisons answer different questions and cannot establish a universal advantage. The announcement provides neither sample sizes nor uncertainty intervals sufficient to reproduce the gains. Searches for independent replication identified no fresh verification of the acquisition-cost result. [2]

Digiday’s October 1 reporting supplies the strongest operational warning. Accuracast said a campaign generated form submissions while ChatGPT still showed zero conversions two weeks later. It also reported roughly 100 platform clicks against 20 in its own tracking, without an explanation from support at that time. Executives at AdRoll, Jellyfish and another agency reported no click discrepancies, however. These are attributed agency accounts rather than an independent audit of platform records. [3]

Some apparent problems need qualification. Web Guide Partner’s reported conversion delay of 24–36 hours falls within OpenAI’s currently documented 24–48-hour processing window; it does not itself demonstrate incorrect counting. OpenAI also explains that clicks and analytics sessions can differ because of consent, browser blocking, redirects or tracking configuration. Those explanations have not been demonstrated for Accuracast’s campaign, and ordinary processing latency does not explain its reported two-week absence of conversions. [3] [4]

Reporting settings also affect comparisons. OpenAI’s documentation says conversion totals can include eligible purchases following an impression without a click, depending on the selected attribution windows. Editorial inference: a credible cross-platform comparison must disclose and align those rules before treating cheaper attributed acquisitions as better advertising. Expanded integrations may ease implementation, but they cannot by themselves reconcile disputed counts or determine how many customers would have purchased anyway. [2] [4]

An academic audit raises a different concern: who bears the advertising exposure. Emma Lurie and colleagues at Pennsylvania and Haverford used 91 simulated accounts with geographic signals associated with income and racial groups. Lower-income accounts were more likely to receive ads, with no detectable race association. Sponsored content remained visually separate, and excluded sensitive categories produced near-zero ad rates. The early evidence therefore flags unequal exposure while also showing some safeguards functioning. [5]

The audit’s limits are substantial. Only 48 accounts received advertisements during its main March window, and collection abruptly declined, possibly because synthetic behavior was detected. Geographic signals are proxies, and automated querying differs from ordinary use. The arXiv record states that the paper is forthcoming in AAAI/ACM AIES 2026; hosting there alone neither establishes nor rules out peer review. Its early observations cannot establish intentional discrimination or describe October’s planned format. [5] [6]

Privacy scrutiny concerns information flowing back from advertiser websites. Researcher Buchodi reported that an identifier called __obi accompanied requests to OpenAI, including page or conversion information. Retained URL paths could reveal sensitive interests even when query strings were removed. Testing covered Chrome on Android, with incomplete session coverage and desktop Chrome untested. Crucially, the researcher did not observe OpenAI joining received events to accounts on its servers; the design suggests that capability without proving its downstream use. [7]

Tom’s Guide emphasized that those observations do not show advertisers receiving private conversations. Its coverage relies on Buchodi’s investigation rather than providing a separate technical replication. Editorial inference: keeping chats from advertisers and limiting incoming website tracking are distinct privacy questions. Evaluating the latter requires clearer evidence about consent, retention and account linkage. Neither the observed identifier nor assurances about conversation privacy settle the full boundary between an assistant and its advertising infrastructure. [9]

There is limited evidence supporting answer independence. Nile’s secondary analysis found advertiser names in about 8% of sampled answers and an average difference of minus 0.3 percentage points across 91 matched brand-and-prompt pairs with and without ads. It reuses the academic dataset, counts mentions rather than endorsements and is observational. Its causal claims exceed that design. The cautious editorial verdict would improve with fresh answer audits, reproducible sales experiments, reconciled reporting and independent tests of visual disclosures and data boundaries. [8] [2] [4] [7]