China publishes a five-part GEO standard with a 30-run sampling floor

An AI visibility report built on five runs of each prompt may be reading noise instead of movement. Search Engine Journal's Marcus Pentzek reports that the China Association of Commercial Advertising has published a five-part voluntary standard that sets a floor of 30 runs per prompt at the finest level a vendor reports on, and backs it with a confidence table taken from the association's own 2025 industry report.

The series was issued through the China National Institute of Standardization's group standards platform under codes T/CAACCHINA 001 through 005, between 9 and 14 September 2026. Group standards (the "T/" prefix) are voluntary and not law: Pentzek writes that their force comes from brands citing them in contracts and regulators using them as a reference. Publicis Group is listed as a drafter of parts 1 and 5, so a Western agency network has already put its name to the principles in China.

What the five parts cover

Part 1 defines generative engine optimization and separates it from search engine optimization and answer engine optimization. Part 2 lays out six vendor types, five assessment dimensions and an A to D grade with a "one-vote veto" for serious failures. Part 3 defines pricing, metrics and sampling. Part 4 rates sources and sets a correction duty on wrong answers in regulated industries. Part 5 covers audit trails, banned tactics and contract requirements.

The sampling floor is the clearest practical ask. Pentzek quotes the appendix table from part 3: at 5 runs per prompt the standard deviation of measured visibility sits at plus or minus 15 to 24 points, at 10 runs it falls to plus or minus 10 to 16, and at 30 runs it falls to plus or minus 6 to 8. These are one standard deviation values, so a 95 percent range is about twice as wide, and even the 30-run floor leaves a plus or minus 18 point window at 50 percent true visibility. At 5 runs the score can only move in 20-point jumps, and a brand that truly appears in half of all answers will land at or below 20 percent in about one report in five, with nothing in the real world changing.

Where it draws the line between optimization and manipulation

Part 3 also defines core metrics with formulas, so two vendors using the word "visibility" measure the same thing: an AI visibility rate counted only on category prompts (never on branded ones), a share of visibility against named competitors, an average first-mention position, a sentiment non-negative rate and a source citation rate. Pentzek writes that if a model is used to judge sentiment, the vendor must disclose which model, which prompts and which rules, and keep a human reviewer on the sentiment calls.

Part 5 names tactics vendors may not use: fabricated content, manipulated rankings and the "AI poisoning" pattern a China Central Television consumer rights broadcast exposed on 15 March, when an insider planted a fictional smart wristband into AI answers using software-generated reviews. The drafting committee held its first full meeting on 29 April, so the industry had about six months from the scandal to published rules.

Pentzek flags one caveat on the appendix table: the confidence numbers do not match textbook binomial maths (the 20 and 80 percent columns should be identical and are not). The 2025 report behind the table was co-authored by the same working committee that wrote the standard.

An AI visibility dashboard built on fewer than 30 runs per prompt per day is the first thing a buyer can check against this standard, and so is a contract that promises guaranteed positions in AI answers.

Source

China Wrote The First GEO Rulebook After A TV Scandal, Search Engine Journal

T/CAACCHINA 001-2026 and the four following parts, China's group standards platform (ttbz.org.cn)