More than half of the eligible one-day view-through conversions that OpenAI measured after a ChatGPT ad happened in the first hour, even when the user never clicked the ad. That 52.7% figure sits deep in OpenAI's new measurement note. It explains why the company's latest ad format matters: a visual promotion shown while someone is making an image can feed an optimization system after the person leaves ChatGPT and buys elsewhere.
OpenAI announced on October 5 that it will start testing visual ads during image generation later this month in the United States. The ad may depict a product, show how it is used, or present an experience associated with it. OpenAI says the sponsored image will be labeled and kept separate from the image ChatGPT creates. The company also says advertising does not alter ChatGPT's answer.
The placement is easy to picture. Ask ChatGPT to render a room, an outfit, or an event concept, and a separate commercial image can appear near the result. What deserves closer attention is what happens after the impression. OpenAI now lets eligible view-through conversions help determine future ad delivery, turning a moment of visual creation into a measurable step in an advertiser's sales path, according to its measurement note. That is a larger product change than replacing a text card with a picture.
An ad after the image, a signal after the sale
ChatGPT Ads already had image assets inside standard units. OpenAI's advertiser guide says those units can contain an advertiser name, title, copy, landing page, and creative image below a response. The new format is tied specifically to image generation, where the surrounding task is visual and the ad can answer the same prompt context with commercial material. OpenAI has not published a screenshot specification, size requirement, or frequency cap for this test.
The announcement draws a firm interface boundary: the ad remains outside the generated image. That distinction matters when a user may be asking for a product mock-up, mood board, or shopping idea. OpenAI's ad policies also prohibit ads that imitate ChatGPT's interface or voice closely enough to be mistaken for part of the product. The test will show whether a label and visual separation are enough when the response and the ad share a subject.
OpenAI describes ChatGPT as reaching 1.2 billion people each week, a company-reported figure that gives the format unusual potential scale. The first visual test is narrower: one country, an initial group of advertisers, and a start later in October. The launch post does not name those advertisers, list eligible product categories for the format, or say how many users will see it.
The feedback loop extends beyond the click
A click is no longer required for a purchase to influence campaign optimization. OpenAI's measurement documentation says advertisers can send purchase, lead, install, and sign-up events through a pixel or Conversions API. New integrations with Hightouch, Tealium, and LiveRamp connect those events to systems advertisers may already use, while a longer list of attribution companies can support reporting and click measurement.
For reporting, an advertiser can select a click-through attribution window and add a one-day view-through window. OpenAI says its advanced conversion optimization can learn from eligible conversions on either path. A purchase after an impression can therefore shape which people receive later ads, even if the first person did not open the advertiser's site from ChatGPT. The company says these settings change attribution reports but do not change campaign optimization rules themselves in its measurement documentation.
The 52.7% figure needs a careful reading. It covers eligible one-day view-through conversions in an early analysis of global campaigns optimized for deeper sales outcomes, and measures how quickly those conversions followed a matched impression. It does not show that the ad caused every purchase. OpenAI is separately working with Haus, Measured, and WorkMagic on geographic experiments intended to estimate causation, which the company describes as an early effort in the same measurement note.
OpenAI also published partner-reported results: DV Rockerbox measured WeightWatchers' attributed acquisition cost at 15.3% below its blended paid-search benchmark, while WorkMagic estimated 2.3 times more incremental orders for Dose than last-click attribution captured. These are named examples rather than a platform-wide benchmark. The measurement post does not provide campaign spend, sample sizes, or full study methods, so the figures cannot establish typical performance for the new visual format.
The buying system already has the machinery to act on these signals. OpenAI's Ads basics guide describes impression and valid-click billing, conversion-optimized campaigns, and a relevance-weighted second-price auction. It says advertisers can bid for views or clicks, with conversion campaigns optimized toward a supported event. Pairing that auction with view-through purchase data gives OpenAI a route to value an impression that never produced a visit.
What users can turn off
Ads may appear on ChatGPT's Free and Go plans. Plus, Pro, Business, Enterprise, and Edu accounts remain ad-free, according to OpenAI's user FAQ. Free users can also choose an ad-free mode with lower message limits and fewer tools. Image generation is one of the tools that may be unavailable in that mode, so the free route to the task used for this visual-ad test may require accepting the ad-supported experience. Availability varies by region.
Turning off personalization does not remove ads. It stops ChatGPT from using other threads, ad history, and broader topic signals, but the current conversation can still determine what appears. If personalization and memory are enabled, past chats and saved memories may contribute to ad selection. Temporary Chats show no ads, and personalized ads are not initially available in the European Economic Area or Switzerland, the same FAQ says.
OpenAI says advertisers do not receive chats, memories, names, email addresses, precise locations, IP addresses, or sensitive personal information. Advertisers receive aggregated performance reporting, while conversion events can be sent into OpenAI's measurement system through advertiser tools. Those two data flows should not be collapsed into one claim: conversation text stays unavailable to the advertiser under the published user policy, while outcome data still informs delivery and reporting.
A user can hide or report an ad, inspect why it appeared, turn off personalization, and clear ad data from Settings. OpenAI says cleared ad data stops being used for selection immediately but may remain on its servers for up to 30 days. Ad interactions do not enter ChatGPT memory by themselves. Sharing an ad with ChatGPT through the Ask ChatGPT control creates a conversation that may be referenced later if chat-history reference is enabled, according to the user FAQ.
Safeguards meet a visual placement
OpenAI's placement rules bar ads beside conversations involving personal health, mental health, politics, emotionally reliant use, and other sensitive contexts. The content rules are different from the placement rules. Approved health, financial, or legal advertisers can run in limited circumstances in the United States, but their ads still cannot sit beside a sensitive user conversation. Political advertising is prohibited.
The company uses automated review with human oversight for advertisers, creative material, landing pages, and placement. It is also developing evaluation pilots with DoubleVerify and Integral Ad Science so those companies can examine brand-safety controls in a test environment without seeing private conversations. OpenAI has described those efforts as pilots; it has not published their findings in the October announcement.
Image-generation placement gives those rules a direct test. A harmless prompt can produce sensitive-looking material, and an ordinary shopping prompt can carry personal context that a keyword list would miss. OpenAI says its guardrails assess the conversation for advertising suitability. The published material does not explain whether the generated image itself is evaluated before the sponsored image is selected, or whether an ad is chosen before generation finishes.
The first useful evidence will come from the boundary OpenAI has promised: a clearly labeled ad, separate from the generated image, shown only in suitable conversations. Watch whether users can reliably tell the two images apart, whether reports expose bad contextual matches, and whether OpenAI publishes results from its independent evaluation pilots. The format begins as a US test later this month. Its harder test is whether a no-click optimization loop can remain legible to the person whose image prompt started it.