China Built Standards for GEO Before the West Agreed It Works

The most quoted statistic in generative engine optimization measures something narrower than its sellers claim. That doesn’t mean GEO is fake. It means the pitch has outrun the evidence.

In late 2023, a group at Princeton published a paper called GEO: Generative Engine Optimization. It built a benchmark of 10,000 queries, rewrote the same source pages nine different ways, and measured how much of each generated answer could be traced back to each version. It was accepted at KDD 2024.

The headline result has since become the industry’s founding number: visibility up by as much as 40 percent. Two rewrites did most of the work — adding statistics and adding quotations. Keyword stuffing, the tactic SEO spent fifteen years perfecting, scored below doing nothing at all.

That is a real result, reproducible, and worth knowing. It is also considerably narrower than the way it gets quoted.

What the experiment actually held constant

The retrieval set was fixed. The page had already been fetched and was already in the model’s context window. What the rewrite changed was how much of that page got used in the answer and whether it was named.

So the 40 percent answers one question: given that your page was retrieved, how do you get quoted from it? It says nothing about the harder problem — how your page gets retrieved in the first place.

A critical survey published in July 2026 read 45 GEO studies, the Princeton paper included, and stated the limit plainly: no reviewed technique shows a stable causal effect on organic discoverability that holds across engines and over time. A replication that September applied paired, volume-controlled quotation, statistics and citation edits and found no positive pooled effect at all — quotations at −0.325 percentage points, statistics at −0.276, cite-sources at −0.793.

There is a second twist almost no roundup reports. In the original paper, the same tactics move in opposite directions depending on where a page already ranks: roughly +100 percent at position five, roughly −23 percent at position one. Applying quotation and statistics edits to a page that already ranks first can reduce its visibility, because the engine already trusts it and the additions dilute a signal that was working.

The market grew anyway

None of this stopped the budget from arriving.

Frost & Sullivan, working under the China Advertising Association, estimates China’s GEO market at roughly 700 million yuan in 2024, 5.7 billion in 2025, around 13.7 billion in 2026, and 48.8 billion by 2029. Global figures in the same study run from 47.5 billion yuan in 2025 to 427.9 billion by 2029.

Other estimates differ by orders of magnitude — one research house puts China’s 2025 GEO market at 250 million yuan on a narrow definition of service provider revenue, another puts 2026 at 28.6 billion on a broad industry basis. The spread is a signal in itself: nobody has settled what counts. Treat any single figure as a claim with a definition attached, not a market fact.

What can be said with confidence is directional. Budget previously earmarked for SEO, content marketing and PR is being reallocated, and an entire supply side has appeared. More than two thousand firms in China describe themselves as GEO providers, and by the industry’s own accounting, more than 95 percent of them were SEO or advertising agencies before.

The governance line arrived fast

Whatever the measurement problems, the compliance apparatus is no longer theoretical.

On March 14, 2026, China’s first GEO industry self-discipline convention was signed. The next day, CCTV’s consumer rights programme exposed a grey-market practice known as “AI poisoning” — paying to have fabricated products recommended as top choices by large models. In April, a coalition of more than 40 institutions including the National Advertising Institute and Xinhua’s state key laboratory published a Responsible GEO Governance Initiative. In June, a GEO Red Book set out the boundary lines and risk categories.

Since then, group standards have landed in sequence: a commerce-services specification in July, a trusted-communication standard in August, five standards from the China Business Advertisers Association in September, and three from the China Advertising Association in October. A second GEO industry forum is scheduled for Shenzhen on November 21–22, with 30-plus speakers, framed around one promise: let good brands be correctly recommended by AI.

That sequence matters for anyone buying the service. The industry’s own standards body now cautions that effect changes should not be attributed simply to the GEO service itself — a caveat written into the standard text, not a critic’s aside.

The case that gets cited most

The most-quoted proof point in Asia is a Korean company called Sunjin, which works in animal feed, pig farming and meat processing.

At the OAC Regional Communications Forum in Ho Chi Minh City in early November, Ashley Meejeen Lee of Hahm Partner described the engagement. An initial GEO analysis put Sunjin’s presence rate at 5 percent, measured against a fixed question set. Reviewing existing media coverage, the consultancy found the company was not short of information. Its press releases led with productivity data and event results, without explaining Sunjin’s role in the value chain or how it works with farmers.

The revision was structural rather than volumetric. Releases were rewritten to name the industry the company operates in, connect each fact to the wider business, and attach units and benchmarks so the numbers meant something. Certifications were clarified by issuer and evaluation method. Annual results were expanded to discuss growth drivers, value-chain integration and the role of digital operations.

Sunjin issued 13 press releases, which generated 180 media placements. Company-provided measurement showed AI presence rising from 5 percent to 20 percent after three months.

Then Lee added the sentence that matters most, and which most retellings drop: the improvement could not be attributed solely to the change in press-release content, because other factors could have been at play, and any comparison needed consistent questions and measurement conditions across both time points.

Her actual lesson was about information architecture, not optimization tricks: companies do not necessarily need to produce more information. They need to present existing information clearly, consistently, with evidence, in a form that can be understood.

What a brand should actually do

The honest position is that GEO is real but oversold, and the two facts coexist.

Separate retrieval from citation in your own reporting. These are different problems with different fixes. If your pages are not crawled, indexed or reachable without JavaScript, no amount of quotation-adding will help — the binding constraint sits upstream of the tactic. Fix the plumbing first, then optimize the answer.

Stop measuring with screenshots. Ask the same question repeatedly and you will get different source sets; published audits find day-to-day overlap between cited sources in the 0.34 to 0.42 range. Pick twenty questions in a buyer’s actual language, paraphrase them three ways, run each seven or eight times, and keep a control set. A single citation screenshot is an anecdote, not a measurement.

Prioritize third-party corroboration over owned-page polish. Commercial-intent answers lean heavily on sources the model already treats as authorities, and a brand cannot certify itself. The Princeton levers work on pages that are already inside the retrieved set; getting into that set is largely an off-site problem.

Watch rank dependence before you touch a winning page. If a page already ranks first for a query, applying quotation and statistics edits may reduce its visibility. The tactics that lift a mid-ranked page can dilute a top-ranked one.

Track accuracy, not just mentions. A brand can be cited alongside claims it never made, and 2026 research on adversarial benchmarks found standard safety filters cut that success rate by no more than 5.7 percent. Being mentioned and being described correctly are separate KPIs, and the second one is the one that protects you.

Read the standard before you sign the contract. China’s GEO standards now include an explicit caution against attributing effect to the service itself. If a vendor’s proposal promises a visibility percentage without a question set, a sampling protocol and a control, it is not a measurement — it is a forecast.

The uncomfortable summary is that the discipline with the clearest commercial case has the weakest evidence base, and the industry knows it well enough to have written the caveat into its own standards. That is not a reason to skip GEO. It is a reason to buy the part that can be verified and refuse to pay for the part that cannot.