From GEO to GEM: why being cited by AI is no longer enough, and what changes with ads inside ChatGPT
Since August 18, 2026, ChatGPT has shown ads in Italy and across Europe, and Google is testing ads inside AI Mode. An AI answer now has two layers, what the model says about a brand and what the brand pays to show, and the first decides the second. In China this work already has a name, GEM, and three industry standards.
In brief
- Since August 18, 2026, ChatGPT has shown ads in Italy and across Europe, on its free plans. In under 200 days, OpenAI’s advertising reached $1 billion in annualized revenue. Google is testing ads inside AI Mode’s recommendation lists.
- An AI answer now has two layers: what the model says about a brand and what the brand pays to show. Google picks ads based partly on the content of the answer itself: the first layer decides the second.
- GEM, Generative Experience Marketing, is the name circulating in Chinese industry documents for the work that holds the two layers together. In China, three associations published three GEO standards in three months, with metrics and banned practices.
For a year there was only one question: does my brand show up in the answers from ChatGPT, Gemini, and Perplexity? Since August 2026 there is a second one: what sits next to the answer, and who paid for it.
OpenAI sells ads inside ChatGPT, in Italy and across Europe, and Google is testing ad formats inside AI Mode. The work of getting cited by AI, GEO, is all still there. But it now shares the screen with advertising, at the very moment a person compares and chooses. In China, where the phenomenon is further along, this combination already has a name: GEM, Generative Experience Marketing.
Index
In six months, advertising moved into the answers
On August 18, 2026, OpenAI brought ChatGPT ads to 31 European markets, Italy included. They appear only on the Free and Go plans, the free or nearly free tiers, and OpenAI writes that “ads are always clearly labeled and separate from ChatGPT’s answers, and advertising does not influence the answers.”
They are bought on a CPM or CPC basis, with geographic targeting, custom audiences, a pixel, and an API for measuring conversions, and since August 31 on a self-serve basis through the Ads Manager.
That same August 31, OpenAI reported $1 billion in annualized revenue less than 200 days after launch, with tens of thousands of advertisers in more than 40 countries and over a billion people using ChatGPT every week.
The detail that matters for marketers is a different one: the system “uses the context of the ongoing conversation to show a relevant ad.” The ad lands on the question the person just asked, inside the answer they are reading.
Google took two routes. The first is AI Max for Search campaigns, in beta since May 2025: Google reports 14% more conversions at a similar cost per conversion, and 27% more for advertisers who were still using only exact and phrase match (internal data). Its job is to reach the new searches that AI answers generate.
Ads above and below AI Overviews run in every market; ads inside AI Overviews are for now available only in English in twelve countries, including the United States, Canada, India, and Australia. Italy is not on the list yet.
The second route is newer. On May 20, 2026, Google announced two formats being tested inside AI Mode, built with Gemini. With Conversational Discovery ads, the ad copy is generated for the person’s specific question. With Highlighted Answers, when AI Mode compiles a list of recommendations, an ad can be placed on that list.
In both cases Gemini writes an independent explainer about the product next to the advertiser’s material, and the ad stays labeled Sponsored.
Two ways into an answer: being recommended or being shown
Until now a brand had a single way to appear in an AI answer: getting cited by the model. That is the work of GEO, which acts on sources, structure, and the clarity of information so that models understand and cite a brand. Now there is a second way: paying to be shown next to the answer, or inside it.
The two are not equivalent. OpenAI keeps the ad separate from the answer and states that it does not influence it: buying ads on ChatGPT does not change a single word of what ChatGPT says about a brand.
Google goes further, because with Highlighted Answers the ad enters the list the AI has compiled, with an explainer written by Gemini. For the person reading, it is one list.
The difference that remains is this: the organic route is what the model believes about a brand; the paid route is what the brand pays to display.
The decision takes shape inside the conversation
In a classic search, a person opens ten links and compares them alone. In a conversation with AI, the model does the comparing, and the person reads one answer. Google reports that 75% of AI Mode users make faster, more confident decisions (internal survey). Anyone who wants to be present in that moment has to watch both layers.
The mechanism that ties them together is written in Google’s own documentation: ads inside AI Overviews are selected by considering “both the user query and the content of the AI Overview.” The ad, in other words, is matched to what the AI has just written.
If the answer describes the brand as expensive, dated, or suited to a different need, the ad arrives under that description, or does not arrive at all. The first layer decides the second.
This is the core idea of GEM. Being visible is not enough: what matters is what the AI says about the brand at the moment the person compares and chooses, and that the ad is consistent with that description.
A brand that pays to appear next to an answer that advises against it is paying to confirm a lost comparison. A brand that has fixed the sources models learn its description from, and then buys the ad, shows up twice in the same answer with the same message. If the AI does not cite a brand, advertising does not replace that for now: it shows the brand next to a comparison it was never part of.
The signal from China: standards, metrics, and a closing window
The term GEM was born in China, where generative AI is more widely used: the 57th CNNIC report counts 602 million generative AI users at the end of 2025, 42.8% of the population, up 25 points in a year. And where the industry has given itself written rules.
In three months, three different associations published GEO standards: the China Advertising Association of Commerce published five standards on September 14, 2026 covering definitions, providers, pricing and measurement, trusted sources, and security; a standard from the China Association of Press Technicians, drafted with Xinhua, has been in force since August 12, 2026; a third comes from the China General Chamber of Commerce.
The Xinhua standard is the most useful to read, because it names the bad practices: corpus poisoning (satellite sites, mass low-quality publishing, hidden instructions in pages), manufactured consensus (the same content on many platforms to make it look like a majority view), answer dominance (one party taking up an abnormal share of the answers on a topic).
It requires three separate repositories, verified facts, sourced opinions, and marketing copy, and bars providers from promising “guaranteed first” or “certain citation by the model.” It also sets the metrics: mention rate in answers, accurate mentions, sentiment, accuracy, source citations, and cited pages. The automotive sector is listed as high risk, because of comparative advertising and pricing.
The market explains the urgency. The research firm iAnalysis estimates China’s GEO market at 4.8 billion yuan in 2026 and 137.6 billion in 2030, roughly one tenth of the country’s online advertising flowing to AI apps by that date.
The same study warns that the window with the highest returns “will not last beyond a year,” because advertising inside AI apps will drive up the cost of getting into the answers. That is exactly what OpenAI and Google have just set in motion in Europe.
What a brand does, and in what order
- Measure how AI describes the brand today, question by question and platform by platform: whether it appears, in what position, in what tone, with what errors, from which sources. The metrics in the Chinese standard are a good starting list.
- Make the sources clear and complete, the ones models learn from: website, product pages, structured data, third-party pages the models cite. This is the work that changes the first layer, the only one advertising cannot touch.
- Only then buy advertising, where the recommendation does not reach or where competitors are already on the list. In Italy today that means ChatGPT Ads; ads inside Google’s AI Overviews will follow. And measure the ad against the same description: what does the AI say next to my ad?
The order matters for two reasons. The first is technical: Google matches the ad to the content of the answer, so a wrong description of the brand also hurts the ad’s performance. The second is economic: the organic route costs work and time, while the paid route will keep getting more expensive.
The same applies to the AI agents that are starting to shop on people’s behalf: an agent comparing three offers reads the same sources and the same ads. GEM is the name for doing this work together, with one message across both layers of the answer.
Frequently asked questions
It is the work that brings together the two ways into an AI answer: being recommended by the model (GEO) and being shown through a paid ad. The name has been circulating in Chinese industry documents since 2026. The idea is that the decision takes shape inside the conversation, so the brand’s organic description and the ad have to say the same thing.
Yes, since August 18, 2026, in 31 European markets including Italy. Ads appear only on the Free and Go plans, are labeled, and are kept separate from the answers. Advertisers buy on a CPM or CPC basis through OpenAI’s Ads Manager, with geographic targeting, custom audiences, a pixel, and a conversions API.
No. OpenAI states that advertising does not influence ChatGPT’s answers. The ad appears next to the answer, but the model describes the brand based on the sources it has read. Changing that description takes clear, complete, consistent sources: that is the work of GEO, and it comes before buying ads.
