China published the first GEO standard: what it demands from anyone selling AI visibility

On September 14, 2026, the China Association of Commercial Advertising (CAAC), a Chinese trade body for commercial advertising, published five GEO standards on the national group standards platform. GEO is the work of getting a brand mentioned and recommended in AI answers. The documents, numbered T/CAACCHINA 001 to 005—2026, define what counts as GEO, how to measure it, how to price it and what is off-limits. Search Engine Journal called the set the industry’s first rulebook.

The standard is voluntary and was written by the people who sell GEO. Even so, it turns into written rules what buyers anywhere can already demand in a contract. Its weak spot is measurement: the minimum sample it sets is below what the data suggests even for a first diagnosis.

What pushed China to write GEO rules

The trigger was television. On March 15, 2026, World Consumer Rights Day, state broadcaster CCTV’s 3·15 show ran a test with a smart bracelet that doesn’t exist, the Apollo-9. According to China Daily, GEO software bought on an e-commerce platform generated more than a dozen promotional articles with invented features and posted them to self-publishing accounts. After 11 more fabricated articles, two AI assistants listed the device among their top bracelet picks. The report doesn’t say which ones.

The response came on two fronts. According to the Global Times, SAMR, China’s market regulator, had already named AI-generated advertising an enforcement challenge for 2026 back in January. On the industry side, the drafting committee first met in Beijing on April 29, according to Search Engine Journal.

What the standard defines

The numbers below come from Search Engine Journal, which linked and described all five parts. The official platform lists the documents but doesn’t display the full text.

• Five parts — terms and definitions (separating GEO from SEO), vendor evaluation with A-to-D grades and a veto, pricing and measurement, source quality, and process compliance.
• Who wrote it — vendors and platforms: Mininglamp led the pricing and measurement part, Publicis is listed on two parts, and Zhihu, Sohu, Weibo’s operator and 360 took part.
• Metrics — AI visibility rate, share of visibility against competitors, average position, non-negative sentiment rate and source citation rate.
• Minimum sample — 30 runs at the most granular reporting level, for example per intent, per engine and per day.
• Contracts — in hybrid pricing, the fixed portion must be at least 30% of the contract value; error corrections must start within 72 hours.
• Prohibited — fake content, corpus poisoning, prompt injection, forged reviews and sources, smearing competitors. Promising “absolute control” of model output or “100% dominance” is also banned.
• Legal force — none. It’s a group standard, with no fines and no effective date given by the sources.
• Outside China — it doesn’t apply. There’s no mention of other markets.

Why it matters to anyone hiring GEO

China’s standard answers in writing the questions I listed in my guide to evaluating a GEO specialist: how many answers do you collect, what can’t you promise, how do you handle manipulation. The difference is that there they became contract clauses.

That gives buyers a ready-made reference. In Brazil, CRO Brasil summed it up as “we can just copy it” and listed the clauses: minimum sample, a list of what can’t be promised, a ban on black-hat GEO, a fixed fee for legitimate work and an error-correction plan.

The limits are in the sources too. Vendors wrote most of the standard, which may protect the market as much as the buyer, as both Search Engine Journal and CRO Brasil point out. It’s voluntary, so it only binds when it goes into a contract. And part of it, according to CRO Brasil, is traditional SEO under a new name.

What the data says about the minimum sample

Thirty runs per slice is too few to claim a brand’s presence. In FA-MET-01, “How to Measure What an AI Recommends”, a methodological study First Answer published in April 2026 with three experiments and 7,000 runs in the ChatGPT interface, even a first diagnosis, which only shows whether a market is concentrated and who leads it, takes 100 to 200 runs, with a ±7 to 10 point margin. Measuring each brand’s presence within ±3 points takes 1,000.

The standard itself hints at the problem. According to Search Engine Journal, the margin table in the Part 3 annex doesn’t follow classical binomial statistics, and a 95% interval would be roughly double what it shows.

The access environment is also missing. The coverage doesn’t say whether the standard requires stating if measurement ran through the API or the interface consumers use. The second study in the series, FA-MET-02, “API or Interface: That Is the Question”, asked the same question 1,000 times through the ChatGPT API and 1,000 times through the interface: 12 of the 15 main brands shifted significantly between the two environments.

What to ask for in a GEO contract

1. Get the sample in writing: how many runs per question, per AI engine and per period, and the margin of error.
2. Ask for the measurement environment: API or interface, and which AI engines.
3. Attach a list of what can’t be promised, such as a guaranteed top spot or control over the answer.
4. Add a clause against fabricated content, naming the prohibited tactics.
5. Set a correction deadline for when an AI repeats false information about the brand.

A yardstick that still needs scale

China was first to write down what honest GEO looks like, and the playbook works in any market. The number it set for measurement is a starting point, not the finish line.