Ads in ChatGPT and GEO: what can be measured today, and what is just a promise
Ads in ChatGPT are appearing in a growing share of conversations, and the industry is urging optimization for AI answers. We examined which numbers are hard data and how to sensibly split budget and attention between these two paths.
A marketer who wants to be visible in ChatGPT has two paths today. The first is paid advertising, which OpenAI has been developing since February 2026. The second is GEO (generative engine optimization, meaning refining content so that AI assistants mention a brand in their answers). Both paths have their supporters, and both have accumulated plenty of numbers around them. Those numbers need to be approached cautiously, however, because some of them measure something different from what the headlines suggest.
Advertising does not buy placement in the answer
The most important distinction came up during a webinar that Go Fish Digital prepared together with OpenAI. Abhilash Edathil from OpenAI's monetization team showed a conversation about travel planning in which an ad appeared next to the answer, but separately from it. The ad can use the context of the conversation and, with the user's consent, personalization as well. According to Edathil, however, it does not influence the content of the answer itself.
AJ from Go Fish Digital called a recommendation in the answer “earned” visibility and a labeled ad “bought” visibility. So two different brands can appear on one screen, and there is no contradiction in that. For budget planning, this leads to a simple conclusion: a paid test can show an offer at a relevant moment in the conversation, but it will not make the model start recommending the company. A mention in the answer, in turn, is not an ad impression or proof that someone clicked.
How many ads are really visible in conversations
The most talked-about number of recent days comes from Comscore's report for the second quarter of 2026, published on September 22. In June, sponsored ads appeared in 47 percent of tracked ChatGPT conversations about video games and consoles, compared with 3 percent in March. In the toys and hobby category, the share rose from 5 to 24 percent, and in books, music, and video it increased from 9 to 22 percent, after a May drop to 8 percent.
These data, however, have clear limitations. They apply only to desktop activity in the US, come from a panel that users join voluntarily, and count conversations rather than users or impressions. Comscore did not say how many conversations stand behind each month, how it assigns them to categories, or whether a single ad anywhere in a conversation is enough. It is also unclear whether the denominator includes conversations from Plus, Pro, Business, Enterprise, and Edu accounts, which do not show ads at all - ads are only seen by users on Free and Go plans and by logged-out users.
The June jump coincides with the launch on May 5 of the self-serve Ads Manager in beta and the addition of cost-per-click (CPC) billing alongside cost-per-thousand-impressions (CPM) billing. Comscore does not link these facts causally. Other measurements are hard to compare with it: SE Ranking found ads in 25.94 percent of 50,006 commercial queries, but those were written by the research company and collected on a single day, July 23. The ppc.land site reports that on August 31, ads in ChatGPT reached an annual revenue run rate of around $1 billion, and the self-serve Ads Manager was also opened at that time to advertisers from Europe. That information comes from a single trade publisher, not from an OpenAI financial report.
GEO: less magic, more order on the website
Marketing materials still circulate the promise of “30-115 percent higher visibility” after applying GEO techniques. It comes from the first study that popularized the term, and it was measured under experimental conditions. A critical review of studies from 2023-2026, published on arXiv, is much more cautious. Content that the model has already pulled into context may influence the answer, but the authors did not find a technique with a lasting effect across different platforms that would translate into discoverability, clicks, or conversions.
Practitioners at Go Fish Digital are essentially saying the same thing, just in the language of everyday work. Before a company starts seeking mentions on other sites, it should get its own website in order, because that is the only thing it fully controls. Patrick Algrim gave the example of a moving company that handles interstate moves but does not write about it on its website and has no confirmation of it anywhere else. It should not expect an AI system to figure that out on its own. AJ added a technical checkpoint: pages must be accessible to indexing crawlers, not blocked by rules in the robots.txt file or by the content delivery network (CDN). Reviews and PR still matter, but they will not fix an unclear or contradictory description of the company.
Visibility in AI is a pattern, not a ranking position
Model answers are too variable to treat like a position in Google results. AJ suggests evaluating three things: whether the brand appears at all for buyers' questions, whether it is described accurately and up to date, and how often it appears compared with competitors. The method is simple: gather 20-40 questions that customers actually ask, repeat them for a week or two using the same model and settings, and then look for recurring gaps.
Measuring traffic is even harder. Someone may discover a brand in chat and then search for it separately or type the address manually, so in a “last-click” attribution model the impact of AI disappears. Referrals from assistants in Google Analytics 4 show only part of the path, so it is worth comparing them with traffic from brand-name queries, direct visits, and customers' own declarations. The scale is well illustrated by Conductor's aggregate data from November 2025: referral traffic from AI accounted for an average of about 1 percent of site traffic, while traditional organic search accounted for about 25 percent.
An interesting signal comes from Baidu, which on September 16 confirmed a new AI citation report in its webmaster tools. It will show citations, link impressions, clicks, and click-through rate (CTR), calculated, however, as clicks divided by citations rather than by impressions. It will cover only AI answers in regular Baidu results, and the dashboards are empty for now. It is a good lesson: “visibility in AI” means something different depending on who measures it and how.
How to sensibly split the effort today
Our conclusion is that ads in ChatGPT and GEO are not interchangeable and should not compete for a single budget. Advertising can be measured in the traditional way, but its reach depends on the market, the industry, and the user's plan. The effects of GEO currently cannot be reliably translated into revenue, and the work that supports it improves the website for people anyway. The practical order looks like this:
- List buyers' questions and check repeatedly whether the brand appears in the answers, how it is described, and how it compares with competitors.
- Get the pages that are supposed to answer those questions in order: services, terms, differentiators, access for indexing crawlers.
- If paid presence fits the goal, launch a limited ad test with one defined objective and conversion tracking connected.
- Read AI traffic data together with branded and direct traffic, and record every website or campaign change with the date.
Anyone hearing promises today of guaranteed GEO gains or drawing conclusions from a single Comscore chart should first ask what exactly was measured. The safest investment remains a website that clearly and truthfully says what the company does.