China Wrote The First GEO Rulebook After A TV Scandal – Western SEOs Should Take Notes

In a section aired in March 2026, an trade insider sat in entrance of a hidden digicam and invented a sensible wristband that didn’t exist.

Software program generated a primary batch of greater than a dozen promotional articles, and he adopted up with 11 extra: knowledgeable opinions, trade rankings, and person evaluations. When reporters later requested AI fashions for sensible wristband suggestions, a minimum of two listed the fictional system amongst their high picks.

That demonstration aired on China’s most-watched client safety broadcast. It took hours, not months. And nothing concerning the method relied on the AI fashions being Chinese language.

Six months later, China’s promoting trade answered with one thing the Western search trade nonetheless doesn’t have: a five-part commonplace that defines what generative engine optimization (GEO) is, how its outcomes needs to be measured, what a GEO contract could promise, and the place optimization ends and manipulation begins.

I’ve adopted the Chinese language search marketplace for years, and that is the primary time I’d advocate that Western SEOs learn a Chinese language trade commonplace intently. Not as a result of it’s good, however as a result of it solutions questions our trade retains arguing about with out resolving.

Right here’s the way it took place, what’s in it, and which concepts are value borrowing.

From Scandal To Normal In 6 Months

To know why the trade moved this quick, that you must know what “3·15” means in China.

Yearly on March 15, World Client Rights Day, China Central Tv (CCTV) airs a primetime gala constructed on undercover investigations into firms that cheat shoppers. Previous editions have uncovered meals security violations, knowledge harvesting apps, and pretend opinions. Being named on the present can imply regulatory motion, collapsing gross sales, and public apologies inside days.

In 2026, the present took goal at GEO. Reporters described distributors who promised purchasers that, for a price, their merchandise would seem continuously in mainstream AI solutions, even when the content material behind these solutions was false promoting. This system known as it “AI poisoning.”

What adopted moved rapidly:

Why The Trade Wrote Its Personal Guidelines

The usual doesn’t say why it was written in so many phrases, however having watched how Chinese language industries reply to three·15 exposés, I see three motives.

Pre-empting regulators. In China, a 3·15 section is commonly adopted by authorities guidelines. An trade that writes its personal commonplace first will get to form what “compliant” means earlier than a regulator defines it extra strictly.

Saving the label. Professional companies had spent two years promoting “GEO” to manufacturers. After primetime tv linked the time period to poisoning AI, they wanted a transparent public line between themselves and black-hat operators. As you’ll see under, the usual attracts that line explicitly, and it protects legit PR and paid media as a lot because it restricts manipulation.

Giving consumers a yardstick. Earlier than the usual, manufacturers had no shared method to examine GEO distributors, value their providers, or confirm outcomes. Analysis agency iiMedia, as reported by Yicai, had flagged low standardization and a scarcity of unified metrics because the younger trade’s core issues.

None of those three issues is uniquely Chinese language. Western SEOs face the identical mixture of overpromising distributors, incompatible AI visibility metrics, and an undefined line between optimization and manipulation. We simply haven’t had the televised scandal but.

What The Normal Truly Covers

The collection was issued by the China Affiliation of Business Promoting (中国商务广告协会). It’s to not be confused with the separate China Promoting Affiliation. The work was organized by the affiliation’s AI Advertising and marketing Software Working Committee. All 5 elements are listed as present on China’s nationwide group requirements platform, which is operated by the China Nationwide Institute of Standardization.

One level issues for a way you learn every thing under: these are group requirements, marked by the “T/” prefix of their codes. They’re voluntary, written by trade individuals, and never regulation. Their power comes from manufacturers writing them into contracts and from regulators utilizing them as a reference.

The listing of contributors is as fascinating because the content material. The platform names dozens of drafting organizations throughout the 5 elements. Mininglamp, a Hong Kong-listed advertising and marketing know-how firm, led the drafting of Half 3. Content material and search platforms resembling Zhihu, Sohu, Weibo’s operator, and 360 additionally took half.

One title will stand out to Western readers: Publicis Group is listed as a drafter of Components 1 and 5. A serious Western company community has already put its title to those ideas in China.

See additionally: Baidu Now Measures AI Citations: Right here’s What SEOs Ought to (And Shouldn’t) Learn Into It

5 Concepts Western SEOs Ought to Borrow

A lot of the usual is written for the Chinese language market. However 5 of its concepts work wherever AI solutions affect shopping for choices.

1. Measure AI Visibility With Sufficient Samples

The issue each AI visibility software faces is that generative engines don’t give the identical reply twice. Half 3 says this instantly: equivalent questions on the identical system and mannequin can produce completely different solutions, so a single run proves little or no.

The usual’s reply is a minimal of 30 runs on the most interesting stage you report on. Its personal instance: when you report per intent, per engine, per day, meaning 30 runs per intent, per engine, on daily basis. It backs this with a desk of the usual deviation of measured visibility at completely different pattern sizes, taken from a 2025 trade report by the China Affiliation of Business Promoting’s AI Advertising and marketing Software Working Committee and Mininglamp Expertise:

Supply: T/CAACCHINA 003—2026, Appendix A, citing the 2025 China GEO Trade Growth Report. Figures as revealed. They’re one commonplace deviation (roughly 68% confidence); a 95% vary is about twice as broad. The revealed values don’t observe textbook binomial maths: the 20% and 80% columns needs to be equivalent, and the 60- and 100-run figures are tighter than concept predicts. Deal with the course as stable and the precise values as indicative.

Learn the 5-run row once more. With solely 5 runs, scores can solely transfer in 20-point jumps, and a model that really seems in half of all solutions will land at 20% or under roughly one report in 5, with nothing truly altering. In case your month-to-month AI visibility report rests on a handful of runs per immediate, its pattern traces could also be noise. Even the usual’s minimal isn’t exact: at 30 runs and 50% visibility, the 95% vary continues to be about ±18 factors.

The usual additionally defines its core metrics with formulation, so two distributors measuring “visibility” are measuring the identical factor:

• AI visibility charge: The share of solutions that point out or advocate the model, which might be reported for top-1 or top-3 positions.
• Share of visibility: The model’s mentions relative to rivals within the class.
• Common place: The place the model first seems within the reply.
• Sentiment non-negative charge: The share of mentions which can be optimistic or impartial.
• Supply quotation charge: How typically official or goal sources are cited.

The usual additionally warns that share of visibility and common place rely closely on how the class and query set are outlined, in order that they shouldn’t function standalone contract KPIs.

Two particulars are value copying instantly. First, visibility charge is calculated solely on class questions (“greatest electrical automobile for households”), by no means on branded ones (“what do you consider Model X”). Branded prompts nearly all the time point out the model, so mixing them in inflates the quantity. Second, if an AI mannequin judges sentiment, the seller should disclose which mannequin, which prompts, and which guidelines it used, and maintain human evaluation within the loop.

What to do: Ask your AI visibility software or company what number of runs sit behind every quantity, and whether or not branded prompts are counted in visibility. If they will’t reply, deal with their pattern traces with warning.

2. Put A Checklist Of “Issues We Can’t Promise” In Each Contract

Half 5 bans distributors from promising results past technical actuality. It names two examples: “absolute management of mannequin output” and “100% display screen dominance.”

It then goes a step additional. Contracts should state the scope, timeline, deliverables, and acceptance standards, and so they should embody a listing of non-promisable objects as a proper appendix.

Western SEOs will acknowledge the issue. “We’ll get you beneficial by ChatGPT” pitches are all over the place, and few of them say what’s exterior the seller’s management: mannequin updates, retrieval modifications, or the straightforward proven fact that nobody controls what an LLM says.

What to do: Earlier than signing any GEO or AI visibility contract, ask the seller for a written listing of what they can not promise. A critical vendor can have one. A vendor who resists the query has advised you one thing helpful.

3. Construction Pricing So It Doesn’t Reward Playing

Half 3 describes 5 pricing fashions: project-based, by service models, performance-based, system-based (platform licenses or customized improvement), and hybrid. The hybrid guidelines are essentially the most sensible.

In a hybrid deal, the mounted portion protecting content material manufacturing, corpus constructing, and monitoring needs to be a minimum of 30% of the contract worth. Efficiency charges are paid in tiers primarily based on how far targets are met. The contract should additionally outline which engines and settings are measured, the baseline, the targets, and the way mannequin updates or rule modifications are dealt with.

The logic is straightforward to observe when you keep in mind the pretend wristband. When a vendor earns nothing until an AI reply modifications, the most affordable method to change it’s to flood the net with fabricated content material. A assured base price for legit work removes a few of that stress.

What to do: In case you’re negotiating performance-based GEO pricing, agree on the measurement baseline and a plan for mannequin updates earlier than you agree on any bonus.

4. Draw A Clear Line Between Optimization And Manipulation

That is essentially the most transferable thought in the entire commonplace.

The usual defines black-hat GEO by way of particular ways: pretend content material, irregular distribution, corpus poisoning, immediate injection assaults, pretend visitors and opinions, solid sources, and maliciously disparaging rivals. Contracts should comprise a clause prohibiting black-hat GEO, and distributors ought to refuse work aimed toward deceptive customers or evading platform guidelines.

Then comes the essential carve-out. Professional promoting, PR, media placements, and influencer collaborations are explicitly not black-hat GEO, so long as the content material is truthful and lawful and nothing is being manipulated.

That sentence does lots of work. It tells manufacturers that incomes protection, publishing knowledgeable content material, and dealing with creators stay honest methods to affect what AI engines study. It locates the issue in deception and fabrication, not within the act of attempting to be seen.

The usual backs this with a “one-vote veto” in its vendor evaluation. Any supplier concerned in black-hat GEO, corpus poisoning, or comparable manipulation fails the analysis outright, nevertheless robust its different scores.

What to do: Write a definition of black-hat GEO into your company agreements. The usual’s listing of ways, plus its PR carve-out, is an effective beginning draft.

5. Plan For When AI Will get Your Model Incorrect

Half 5 recommends an error-correction mechanism in an annex. When content material a vendor positioned seems to be fallacious, or an AI engine misquotes it, the seller ought to begin corrections inside 72 hours. Meaning updating the content material it controls and serving to the shopper replace its web site, data bases, and different channels.

The identical annex requires an emergency response. If a corpus error causes a serious incident, resembling a fallacious value that strikes a inventory or triggers client complaints, associated GEO exercise ought to pause.

Half 4 additionally units stricter sourcing guidelines for delicate industries resembling healthcare and finance. There, solely whitelisted sources rely, and nameless posts and unverified person content material are dominated out. Western SEOs will acknowledge this as a cousin of Google’s “Your Cash or Your Life” (YMYL) logic.

What to do: Deal with AI misinformation about your model like every other repute incident. Determine upfront who displays AI solutions, who fixes the supply content material, and how briskly.

See additionally: Model Safety In AI Search: How To Audit And Defend Your Model’s Identification

The place To Keep Skeptical

A typical is simply pretty much as good because the incentives of the individuals who wrote it. Earlier than you undertake any of the concepts above, maintain these limits in thoughts.

It was written by the trade it regulates. The drafters are principally GEO distributors, companies, and advertising and marketing know-how firms. Mininglamp led the drafting of Half 3, which covers pricing and measurement, and it co-authored the 2025 trade report with the committee that organized the work. That doesn’t make the content material fallacious, however guidelines written by sellers have a tendency to guard the market as a lot as the customer.

The 30-run rule comes from contained in the room. The usual cites a 2025 trade report by the identical committee and Mininglamp as the premise for its sample-size desk. The underlying precept is primary sampling statistics, so I’d count on impartial assessments to substantiate the course. Nonetheless, deal with the precise numbers as a rule of thumb, not settled science.

It’s voluntary. Nothing forces a vendor to observe a gaggle commonplace. Its actual energy will come from manufacturers making compliance a contract situation, and from regulators treating it as a benchmark of excellent observe. We’ll solely know in a 12 months or two whether or not both occurs.

A few of it’s outdated web optimization with new labels. Half 1’s four-layer framework (technical basis, content material high quality, platform adaptation, model thoughts share) maps intently to technical web optimization, E-E-A-T, and model positioning. The usual even cites Google’s E-E-A-T by title, with a be aware that it hasn’t been validated for generative engines. Western readers gained’t discover a lot new there, and that’s advantageous. The worth lies within the measurement, contract, and compliance guidelines.

It was constructed for a unique ecosystem. The usual was written with DeepSeek, Doubao, Kimi, and different Chinese language engines in thoughts, in a market the place a state broadcaster’s exposé can set off enforcement inside days. The concepts switch to ChatGPT, Gemini, Perplexity, and AI Overviews. The enforcement atmosphere doesn’t.

Does The West Want Its Personal 3·15 Second?

The pretend wristband labored on Chinese language AI fashions. I see no cause to imagine the identical method would fail on Western ones. Retrieval-based solutions all over the place reward quantity, repetition, and sources that look authoritative.

The distinction is visibility. China had a primetime demonstration, a regulator’s response, and an trade that wanted to restore its repute quick. Within the West, comparable manipulation can occur quietly, and no person has but had a powerful sufficient cause to jot down the principles down.

If a Western model ever comes, who would write it? Trade associations are one candidate. The AI platforms are one other, since they’ve essentially the most to lose if customers cease trusting solutions. Notably, China’s effort included content material platforms like Zhihu and Weibo, the sorts of sources AI engines draw from. Might Reddit, evaluation platforms, and main publishers sit at an analogous desk?

Or the push might come from consumers. Manufacturers that collectively demand pattern sizes, non-promise lists, and black-hat clauses of their contracts would create a normal in observe, with out anybody publishing one.

You don’t have to attend for any of that. Each thought on this article can go into your subsequent measurement setup or company contract:

• Run every tracked immediate sufficient instances to belief the quantity, and maintain branded prompts out of visibility charges.
• Ask distributors for a written listing of what they can not promise.
• Hold a set price for legit work in performance-based offers.
• Outline black-hat GEO in writing, with an specific carve-out for sincere PR and content material.
• Have a plan for correcting AI errors about your model inside days, not months.

China wanted a nationwide TV scandal to get there. The remainder of us can merely learn their rulebook.

Extra Assets:

Featured Picture: Pawel Michalowski/Shutterstock