China has written the rules of GEO: what its standards say, and what Europe still lacks
Between August and September 2026, two Chinese associations published six standards on GEO: the banned practices listed and defined one by one, separate repositories for facts and marketing, promises a provider may not make, mandatory indicators. We read them in the original sources: what they say and what applies to a brand outside China too.
In brief
- Between August and September 2026, two Chinese associations published six industry standards on GEO, generative engine optimization. Nothing comparable exists in Europe or the United States.
- The standard drafted with Xinhua lists and defines the bad practices one by one, requires separate repositories for facts, opinions, and marketing, bars providers from promising a “guaranteed first place,” and sets the indicators to measure.
- The context: 602 million people in China use generative AI, 42.8% of the population, and the GEO market is estimated to grow from 4.8 to 137.6 billion yuan between 2026 and 2030.
In Europe, GEO, the work of getting cited by ChatGPT, Gemini, and Google’s AI Overviews, is still a practice without rules: every agency defines it its own way and promises whatever it likes. In China, in the summer of 2026, it became the subject of industry standards, with definitions, prohibitions, and indicators. We read them in the original sources: here is what they say and why they matter for a brand outside China too.
Index
Two associations, six standards, three months
The first is T/CAPT 026-2026, the “Standard for trustworthy information dissemination and information ecosystem governance in GEO”: published on August 11, 2026 and in force since August 12 by the China Association of Press Technicians, drafted by Xinhua, the state news agency, with about thirty organizations including Fudan University, the Beijing University of Posts and Telecommunications, and several publishing groups. Twenty-two pages.
The other five are T/CAACCHINA 001-005-2026 from the China Advertising Association of Commerce: work started in March, public consultation in July, publication on September 14, 2026, with 41 organizations and 75 experts on the drafting team. They cover terms and definitions, provider evaluation, pricing and measurement, trusted corpora, security and compliance. The full texts can be downloaded from the association’s site, which presented them in Shanghai on October 16, 2026. Before all of them, in March, came a lesser-known document: the “basic trustworthiness requirements for GEO services” of the AIIA alliance, led by CAICT, the research institute of the Ministry of Industry, with a first round of assessments and nine providers certified in May.
The context explains the hurry. According to the 57th CNNIC report, at the end of 2025 602 million people in China were using generative AI, 42.8% of the population, 25 points more than a year earlier. And the research firm iAnalysis estimates China’s GEO market at 4.8 billion yuan in 2026 and 137.6 billion in 2030. When a market grows thirtyfold in four years, rules follow on their own. The same context gave rise to GEM, the combination of GEO and advertising inside AI answers, which already has a name in China.
The five signals AI uses to decide whom to cite
The first part of the advertising association’s standards does something no official document had done before: it lists the signals a generative AI uses to decide which sources to cite. There are five: semantic relevance, information gain, entity consistency, brand authority, content quality. Alongside these signals, the standard sets out a four-level model: technical foundation, content quality, platform fit, position in the customer’s mind.
This is not Google’s or OpenAI’s secret formula, which, of course, stays secret. It is a description shared by 41 companies that do this work every day, and it matches what we see in the answers: AI Overviews use five sources per answer, chosen from the pages that answer a precise question with verifiable facts. “Information gain” is the item to underline: a page gets in if it adds something the others do not say.
The Xinhua standard defines GEO the same way: the practice that makes information “easier to understand, index, and cite in a reasonable manner” for generative AI, through structure, semantics, credibility verification, and relevance. Which is, give or take a word, the definition we start from. Among its references the standard cites the 2023 Princeton study that coined the term, Google’s spam policies, OpenAI’s usage policies, the ISO 42001 standard, and the NIST AI risk management framework.
The banned practices, one by one
The most useful part of the Xinhua standard is its vocabulary of things you do not do. Four definitions, each with its own article number:
- Corpus poisoning: deliberately inserting false, misleading, or redundant content into the sources the models read. In practice: satellite sites, mass low-quality publishing, hidden instructions in pages, deceptive metadata.
- Manufactured consensus: publishing the same content, or near-identical content, on many platforms so the model takes it for a fact or for the majority view.
- Answer dominance: taking up an abnormal share of AI answers on a topic with content at scale, reducing the variety of sources.
- Prompt injection: hiding instructions in pages, documents, or metadata that push the model outside its normal behavior.
The table of prohibitions adds fake rankings, certifications, reviews, and expert opinions, hijacking competitors’ brands and traffic, and removing the labels on AI-generated content. One by one, these are the services many European agencies sell today without calling them that.
The standard also asks for something every brand should do on its own: three separate repositories. A repository of facts, with evidence and source classification; a repository of opinions, with author and date; and a repository of marketing expressions, slogans, and claims, which “may not replace the repository of facts as the factual basis.” Every factual statement used in marketing must match the version in the repository of facts. It is an exact description of how a model ends up describing a brand badly: when it reads a slogan and takes it for data.
What a provider may promise and what it must measure
Clause 8.3 is the one that changes the market. A GEO provider may not promise or imply a “guaranteed first place,” “stable top three,” “permanent indexing,” “certain citation by the model,” “officially recommended,” or “exclusive platform partner.” It may only agree on process objectives: the share of statements backed by evidence, the coverage of source classification, how often measurements are taken, correction times, delivery deadlines.
The indicators, in Appendix E, fall into five families. Outcome indicators: mention rate in answers, effective mentions, recall of key strengths, share of positive sentiment, accuracy. Citation indicators: total citations, question coverage, pages cited, activity per page. And indicators on the information environment, which no Western provider measures: share of unproven statements, share of misleading expressions, change in source diversity, signs of manufactured consensus. The standard calls for sampling across several platforms, at several points in time, and with independent sessions, because the same question gives different answers.
There are deadlines too. A high risk, for example false information or an attack on competitors, must be picked up within one hour and contained within four; a medium risk verified within 24 hours; an ordinary correction closed within seven working days. And the automotive sector appears in the table of high-risk situations, for comparative advertising, prices, and the basis of reviews.
What changes for a European brand
The Chinese standards do not apply in Europe, and they do not apply to ChatGPT or Gemini. Three things, however, hold everywhere, because they describe how the models work and not how China works.
First: the list of banned practices is a checklist. If a provider proposes identical content on ten platforms, satellite sites, or made-to-order reviews, in China it would be breaking a written standard; here it only breaks common sense, but the model reacts the same way, and Google and OpenAI have their own spam rules, which the standard cites.
Second: the banned promises are a filter for proposals. Nobody can guarantee a citation by a model, because the model is not theirs. Whoever promises it is selling something they do not control.
Third: the indicators are the right ones. Mention rate, accuracy, sentiment, citations, and source diversity are the same measures we use to track a brand cycle after cycle, with six recorded indicators, and they now have an external reference, written by the people building that market. The real difference between the two worlds comes down to one thing: in China, whoever does GEO must keep a record of what they do and correct it within set deadlines. In Europe, for now, that is up to whoever feels like it.
Frequently asked questions
No. They are voluntary industry standards, issued by Chinese associations for the Chinese market. No obligation applies outside China. What applies everywhere is the technical content: the definitions of bad practices and the indicators used to measure results.
It is the publication of the same content, or near-identical content, on many platforms so that the model mistakes it for a fact or for the majority view. The Chinese standard bans it, together with corpus poisoning through satellite sites and answer dominance, that is, taking up an abnormal share of the answers on a topic.
Only process objectives: how many statements are backed by evidence, how many sources are classified, how often results are measured, how quickly errors are corrected. It cannot promise first place, a certain citation by the model, or permanent indexing, because the final result depends on the model and not on the provider.
