China's ad industry sets 30 runs per prompt for AI visibility
Summary
A Chinese advertising-industry standard published in September requires 30 runs per tracked prompt before an AI visibility figure can be reported, because the same question put to the same model returns different answers each time.
Ask whichever AI visibility tool you buy how many runs sit behind each number it reports.
A Chinese advertising-industry standard published in September sets a minimum of 30 runs per tracked prompt before an AI visibility figure can be reported, per Search Engine Journal’s walkthrough of the five-part series. No Western industry body has published an equivalent, and the rule hands buyers an easy question to put to any tracker they pay for: how many runs sit behind each number on the dashboard?
The series runs from T/CAACCHINA 001 to 005 and was issued by the China Association of Commercial Advertising. It defines generative engine optimization (GEO), the business of getting brands named in AI answers, then sets rules for pricing it, measuring it, auditing it and contracting for it. The five parts were published on China’s national group standards platform in the week of September 9. The “T/” prefix marks a group standard, which is voluntary, written by industry participants, and not law. Its force comes from brands writing it into contracts, which makes it closer to a procurement template than a regulation. Publicis Group is listed as a drafter of two parts.
The standard is a response to a televised demonstration of how cheap AI mentions are to manufacture. On March 15, CCTV’s annual consumer-rights gala showed an industry insider inventing a smart wristband that did not exist. Software generated more than a dozen promotional articles for the product, and the insider followed with 11 more written as expert reviews, industry rankings and user evaluations. When reporters later asked AI models for smart wristband recommendations, at least two listed the fictional device among their top picks. The program called it “AI poisoning”, and nothing in the method depends on the models being Chinese: four AI search tools cited an invented concept within a week in a separate test.
Part 3 covers pricing and measurement, and it starts from the fact that generative engines do not give the same answer twice. Identical questions put to the same model on the same device return different answers, so one run proves very little. The rule is 30 runs at the finest level reported: a vendor reporting per intent, per engine, per day owes 30 runs per intent, per engine, every day.
Appendix A of Part 3 shows what a smaller sample costs in accuracy, using figures from a 2025 industry report by the association’s AI Marketing Application Working Committee and Mininglamp, the Hong Kong-listed marketing technology company that led the drafting of Part 3.
| Runs per question | Error bar |
|---|---|
| 5 | ±15 to ±24 points |
| 10 | ±10 to ±16 points |
| 30 | ±6 to ±8 points |
| 100 | ±2 points |
The ranges cover brands whose true visibility is 20%, 50% and 80%. Those error bars are one standard deviation, around 68% confidence, so a 95% range is roughly twice as wide. A five-run sample carries an error bar wider than most of the differences people act on. At 30 runs the window narrows to around 8 points either side of the true figure, which is tight enough to see a real move and loose enough that a small week-over-week wobble is still noise. Search Engine Journal’s piece notes that the published values do not follow textbook binomial maths, so use the table to pick a sample size rather than as a significance test. An earlier test on this site found the same problem from the other side: AI visibility swings of 20 points can be noise at 10 runs a prompt.
Branded prompts do not belong in the rate
Under the standard, visibility rate is calculated only on category questions, never on branded ones. A prompt that names the brand will almost always return the brand, so mixing branded prompts into the set lifts the rate without telling anyone whether the brand gets recommended when it is not named. A vendor reporting a single blended visibility number should be asked what share of its prompt set names the client.
What to do
- Ask every AI visibility vendor for runs per prompt, per engine, per reporting period. Treat anything under 30 as directional and stop reading week-to-week changes off it.
- Rebuild the visibility rate on category prompts only. Report branded prompt performance separately, under its own name.
- Ask for a written appendix of what the vendor will not promise. The standard’s examples are “absolute control of model output” and “100% screen dominance”.
- On performance-weighted contracts, keep at least 30% of the fee fixed, covering content production, corpus building and monitoring, so the work is not paid for by movement alone.
- Borrow the standard’s correction duty: when placed content turns out to be wrong or an engine misquotes it, the vendor starts corrections within 72 hours. Part 4 adds that in regulated categories such as healthcare and finance, only whitelisted sources count.
None of this waits on a regulator. A buyer can put all five into a brief this week, and until a Western body writes its own, the T/ series is the only published text to cite.