With image generation progress at 62%, an ad card for grocery brand Heirloom Grocery appears below, complete with a product image and a “Learn More” button; this is how TNW described the example interface in the official announcement. On October 5, 2026, OpenAI announced that it would test this visual ad format during the image generation process in ChatGPT: ads will be clearly labeled, “remain separate from the image being created”, and “does not influence the answers ChatGPT provides”, launching in a small-scale test in the US later this month.
For a product reaching well over a billion people each week, announcing that ads will be placed into the few seconds users spend waiting for generation carries far more weight than simply adding a new ad slot. This was not a feature rolled out overnight. As early as February 9, 2026, ChatGPT launched ads in the US. Over the following eight months, ad placements shifted step by step: lingering below answers in February, entering conversations in September via testing Sponsored Agents, and being announced on October 5 for testing during the generation process. The ad card in the image generation interface is the latest step on a continuous path. The real question to unpack is: over these eight months, how exactly did the placement of ads arrive here step by step?
I want to look at this journey from a different angle, viewing it as a product manager’s sequence of decisions: under the constraint of not destroying trust, what to fear first, what to test first, and what signals to gather before taking the next step. A disclaimer upfront: I do not have access to OpenAI’s internal meeting minutes or experimental data; the analysis below is a decision reconstruction deduced from public interfaces, official documentation, and release logs, not insider information.
Official figures show that ChatGPT reaches 1.2 billion people weekly, with third-party estimates putting free users at roughly 95%. Every conversation requires running actual inference on servers, and the revenue from the small fraction of users purchasing premium subscriptions and enterprise plans is nowhere near enough to sustain such enormous expenses. For a product operating at the scale of a billion users to make money, the most natural route is adding ads. Sam Altman publicly stated in 2024 that ads were a last resort (with varying reports on the exact month and occasion), but reality intervened, and the mounting bills forced the product to the threshold of monetization.
Inside the team, however, this path was far from easy, because the core competitive advantage of large language models is trust. Users come with questions expecting objective, neutral judgments. Once they feel answers are tainted with sponsor interests, trust quickly collapses. Competitors targeted this vulnerability: Anthropic pledged in February 2026 that Claude would never run ads, even releasing satirical commercials; in the same space, Perplexity also cited trust concerns in February 2026 to shut down its advertising tests entirely.
How to generate revenue without destroying trust boiled down to three dilemmas facing the product manager: seeing whether users would churn because of ads, whether advertisers would continue paying, and how to place ads closer to user creation without violating the commitment to objective trust.
The first hurdle was whether users would swipe away as soon as they saw an ad. Chat interfaces are inherently personal; suddenly inserting an ad makes many people uncomfortable, or even leads them to close the app.
The team initially adopted a relatively conservative strategy. On January 16, 2026, they published a statement of advertising principles, putting forward five core principles in an attempt to reassure the public. On February 9, they officially launched a small-scale test in the US. The interface design was extremely cautious: ads appeared only at the very bottom of complete answers, located at the edge of the field of view; they were displayed only to logged-in adult users on the free tier and the low-cost Go tier, with no ads shown on Plus, Pro, Business, Enterprise, or Edu; and two control mechanisms were provided, allowing users to disable personalized recommendations, while free users could trade fewer daily message caps for an ad-free experience. The cards were also clearly labeled as sponsored, distinguishing them from text responses.
After about six and a half weeks of experimentation, positive results were released officially. On March 26, 2026, the company officially announced that ads had “no impact on consumer trust metrics, low dismissal rates”. This was solely an official and not independently verified statement, rather than an independent third-party audit. However, an empirical audit of 3,602 ads conducted by researchers at the University of Pennsylvania and Haverford College confirmed that ads indeed remained strictly beneath answer text, accompanied by clear visual separation. The study also found that accounts associated with lower- and middle-income users were more likely to see ads.
With positive signals from this self-reported official narrative in hand, scaling out was the natural next move. The team expanded testing to Canada, Australia, and New Zealand, followed by the UK, Japan, and South Korea. In terms of data, the first step of user retention testing had passed muster.
Once users stayed, whether advertisers would foot the bill became the critical question for the next phase. If cards pinned to the bottom of chats failed to drive genuine purchasing conversions, brands would walk away after trying them once, and monetization would not last.
Over the next four months, the team rounded out infrastructure to match standard performance advertising: building a conversion tracking pixel as reported by Digiday on April 16; launching CPC pricing and a beta self-serve ad manager on May 5; rolling out bulk product catalog uploads on May 12 (allowing retailers to upload up to 1 million items directly); testing larger ad canvases and diverse call-to-action buttons on May 21 (reported by Digiday); adding conversion optimization and third-party tracking in late July according to agency Jolly Good Web; and introducing horizontal product carousel cards on August 6. While official announcements cited DV Rockerbox claiming that WeightWatchers achieved an attributed customer acquisition cost 15.3% below blended paid search benchmarks, and cited Triple Whale claiming that 93% of visitors to Portland Leather from ChatGPT ads were new visitors, these were partner-reported metrics with no unified external benchmarks available for cross-verification.
Revenue figures provided the confidence to double down. On August 31, 2026, the company announced in a business milestone announcement that less than 200 days after launching ads, annualized revenue had already reached $1 billion. Although media outlets cited an anonymous source stating that the ad revenue projections presented to investors targeted $2.5 billion in 2026 and $100 billion by 2030, market research firm eMarketer countered, according to Adweek, predicting that the advertising business would fall 90% below the company’s own five-year revenue forecasts.
Even looking solely at the officially disclosed $1 billion annualized revenue run rate, it is enough to explain why OpenAI designated advertising as “one pillar of OpenAI’s diversified business model” in that very announcement, and quickly expanded self-serve ad buying to 31 European countries, India, Southeast Asia, and over 60 markets including Taiwan.
After completing these two steps, the product manager faced a more nuanced dilemma: although banner ads at the bottom were relatively safe, they remained far removed from scenarios where transactions might actually occur. The placements that unlock true value are often situated right at the moment when a user makes a decision or creates content. How, then, could ads be pushed forward without breaking the earlier commitment to user trust and objectivity?
The breakthrough was hidden within the seams of the company’s own wording. A close reading of the initial five principles and the later Help Center reveals that the promised neutrality was drawn with very strict boundaries, consistently restricted to stating that advertising “does not influence the answers ChatGPT provides”. The original text used “answers” or “responses”, without clarifying whether generated images fell under that protected scope; as for the core conversational interaction and the image generation process itself, the original principles never explicitly declared them off-limits.
This loophole existed, and the team leveraged it to roll out two new features positioned closer to decision-making and creation. The first appeared in the chat flow itself. On September 14, media outlets revealed a click-to-chat ad format (reported by Digiday), and on September 16, the company officially announced testing for Sponsored Agents; once users clicked through, a brand’s agent would continue the conversation right inside the dialogue flow. Advertising was no longer just a static label at the bottom, but a participant in the conversation. Meanwhile, the company also updated its advertising policies, and around the same time, according to The Information, blocked commercial promotions for AI image and audio tools that competed directly with its own features in actual enforcement. The second appeared during image generation. On October 5, the company announced visual ads in image generation, with an example interface displaying an ad card beneath a 62% generation progress bar, with testing scheduled to begin later this month.
At this stage, official messaging continued to emphasize that ads “does not influence the answers”, yet offered no explicit guarantee that the generated images would remain free from ad influence; it stated only that ads “remain separate from the image being created”. Separation is a layout specification, not an algorithmic guarantee of neutrality.
This echoed the warning from former OpenAI researcher Zoe Hitzig after her departure in February 2026, in a guest essay for The New York Times: she voiced concern that future product iterations would struggle to hold the line, because the company was “building an economic engine that creates strong incentives to override its own rules”. Rather than literally breaking the promise to protect answers, the team turned interactions omitted from the fine print into testing grounds for monetization, one by one.
Connecting these three steps reveals a clear evolutionary trajectory: the five public principles served as a defensive protocol, delineating PR red lines to show what the company would not do; the subsequent tactical moves were offensive, conducting incremental trials in areas uncovered by those red lines, and scaling up as soon as positive signals emerged. Advancing commercialization never requires shattering commitments head-on; it merely fills in the spaces that commitments never touched.
Such precedents are common in business history. Google Shopping moved from completely free algorithmic rankings in 2002 to an entirely pay-to-play ad display over a decade by 2012, a shift that the European Court of Justice ultimately upheld with a 2.42 billion euro antitrust fine in 2024. Amazon followed a similar path: testing by consumer groups found that an average of 54% of the first 25 search results were ads, and academic research revealed that paid placements could vault products ranked beyond 100 organically directly onto the front page, with top paid results often being more expensive and of lower quality.
External regulators have also taken notice of this trajectory. On August 31, 2026, the European Commission designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act, requiring it to establish an advertising repository and submit to independent audits no later than January 2027.
This evolution offers a grounded framework for observing AI monetization. Whenever another platform releases a seemingly restrained advertising pledge, rather than being swayed by polished PR rhetoric, one can evaluate three product questions: first, what specific risk is this move defending against, and how is the team probing boundaries; second, are the signals driving expansion merely official and not independently verified narratives, or credible independent third-party audits; and third, which narrow interaction did the written promise actually protect, and what blanks were left behind in emerging interaction scenarios?