On 5 October OpenAI published a post about a new visual ad format for ChatGPT and, further down, an expansion of how advertisers can measure what their ads do. The format is the headline. The measurement section is the part with consequences, because it is where a promise OpenAI has made about privacy meets a pipeline that was built for something else.
The promise, from the ads help page I read the same morning: “We do not share your conversations with ChatGPT with advertisers, and we never sell your data to advertisers.” And: “Advertisers only receive aggregated, non-identifying information about how their ads perform, such as total views or clicks.” Elsewhere on that page, under what is shared with advertisers: “For this early test, advertisers receive aggregated reporting such as views and clicks. We may explore additional measurement insights over time while continuing to protect user privacy.” The page was marked as updated seven days earlier, and it says ad testing started in the US on 9 February 2026.
Every one of those sentences describes data going from OpenAI to the advertiser. That is the direction a privacy promise about chats naturally faces. Measurement, though, mostly runs the other way.
What the pipeline carries
OpenAI’s Conversion Measurement help page, marked “Updated: last month” when I read it, is a plain description of how a purchase on an advertiser’s site gets credited to a ChatGPT ad. When someone clicks, a click reference called oppref is appended to the landing-page URL. The OpenAI Pixel, which the advertiser installs, stores it in a first-party cookie and attaches it to later conversion events. Advertisers can also send events server-side through a Conversions API.
When the click reference is missing, the page offers three fallbacks. Advertisers may send “normalized and hashed contact information” with an event; this is called advanced matching. An automatic version has the Pixel automatically detect “supported customer information from recognizable forms and other sources on your website,” hash it with SHA-256 in the browser, and send that, with the note that “raw customer information is not sent to OpenAI through automatic advanced matching.” And there is modeled measurement: OpenAI “may also use aggregated patterns from observed conversions to estimate attribution for otherwise unattributed advertiser-reported conversion events,” and “reported conversion totals may include modeled conversions where available.”
None of this is exotic. Features like these are common in ad measurement, and the 5 October post adds integrations with Hightouch, Tealium and LiveRamp for sending conversion data in, plus attribution partners such as AppsFlyer, Triple Whale, Adjust, DV Rockerbox, Northbeam, Branch, Singular, Kochava, Airbridge and Tenjin. What matters here is the direction of travel. Identifiers and purchase events flow from advertisers into OpenAI. What flows back out is, per the help page, reporting “designed to show campaign performance rather than individual-level user activity.”
Hashing contact details before sending them is standard, and the help page tells advertisers to send data only “after providing clear and comprehensive information to users” and obtaining required consents. But “hashed” is a technical description, not a privacy one. I covered a version of that point in an earlier essay on chatbot trackers: a hashed email still identifies someone to any party that already holds the same email. The documents I read do not say which OpenAI-side records a hashed address is matched against, whether a matched purchase becomes part of the “ads data” a user can clear, or how long conversion events are kept. The help page’s retention line, up to 30 days after deletion, is about ads history and topics. I could not find the answer to the other two on the pages I read, and the post points to an Ads Blog entry and the help page to developer documentation, neither of which I read, so those answers may live there.
None of that touches the claim that chats stay out of advertisers’ hands, and nothing I read contradicts it. The help page’s “such as views and clicks” is an example list, not an exhaustive one, and the Conversion Measurement page already describes conversion reporting. The ads page calls itself “this early test.”
The three results, read for what they leave out
The post offers three partner-reported figures as “early partner findings,” and I would treat each as a vendor claim, which is what the post itself says by attributing each to the vendor.
DV Rockerbox said WeightWatchers’ “attributed cost per acquisition” on ChatGPT Ads was 15.3% lower than its “blended paid-search benchmark.” One side of that comparison is a single channel’s attributed number. The other is a “blended” paid-search benchmark that the post does not define, so the two numbers may not be computed the same way. Attribution gives a channel credit for conversions that touched it, whether or not the ad caused them, so the gap does not tell you what a ChatGPT ad added. WorkMagic reported “statistically significant lift” for Dose, with 67% of incremental purchases coming from net-new customers. That is the only one of the three that speaks of incrementality, and it gives no lift size, no sample, and no test design. Triple Whale said 93% of Portland Leather’s visitors from ChatGPT Ads were new. Visitors are not purchasers, the figure has no comparison channel in the post, and a cold audience would produce a high new-visitor share for many top-of-funnel sources.
The post is candid about this gap in its own words: “our work on incrementality is still in its early stages.” It names Haus, Measured and WorkMagic as partners “to explore geo-based experiments to help advertisers understand the causal impact of ChatGPT Ads.” That is the right instrument. It is also an admission that the causal question is open, set in the paragraph directly above three numbers presented as illustrating “strong performance.”
An old gap, measured once
There is an old precedent for why experiments matter. In 1995 Leonard Lodish and colleagues published a meta-analysis of 389 real-world TV advertising experiments run on BehaviorScan, a split-cable system in which matched households received different commercials and their purchases were tracked through a consumer panel. Their abstract reports that “increasing advertising budgets in relation to competitors does not increase sales in general,” and that the data “do not show a strong relationship” between standard recall and persuasion copy-test measures and sales effectiveness. A matched-household split-cable experiment was the way to find that out. Click-level attribution answers a cheaper question, which is who touched the ad before buying, and a channel that launches with attribution first and incrementality “in its early stages” is answering the cheaper question first.
None of this means ChatGPT ads don’t work. It means the numbers on offer so far cannot tell an advertiser that. If I were a buyer, I would take the three results as an invitation to run a geo test, and I would read the Ads Blog and developer docs for the attribution window, the modeling disclosure and what happens to a conversion event when a user clears their ads data. If I were a user, I would note that “aggregated” describes what leaves OpenAI, and that most of the new measurement traffic arrives from somewhere else.
References
- OpenAI (2026). Building advertising for the way people use AI. OpenAI, 5 October 2026. The partner results are OpenAI’s and its partners’ own claims.
- OpenAI (2026). Conversion Measurement. OpenAI Help Center; page showed “Updated: last month” when read on 5 October 2026.
- OpenAI (2026). Ads in ChatGPT. OpenAI Help Center; page showed “Updated: 7 days ago” when read on 5 October 2026.
- Lodish, L. M., Abraham, M., Kalmenson, S., Livelsberger, J., Lubetkin, B., Richardson, B., Stevens, M. E. (1995). How T.V. Advertising Works: A Meta-Analysis of 389 Real World Split Cable T.V. Advertising Experiments. Journal of Marketing Research 32(2), May 1995.