China's GEO Compliance Reckoning: The Brands That Lost Their Citations

The rules landed in August. In September, the enforcement started showing up in public.

A month-long investigation published by Hunan Daily on September 10 traced what happened to China's generative engine optimization market since the GEO Trusted Information Dissemination and Information Ecosystem Governance Specification took effect on August 12. The finding is not that the market cleaned up. It is that the market split in two — and the gap between the two paths is widening fast.

For overseas brands, this is the most consequential GEO story of the quarter, because the risky path is the one that looks cheapest and most familiar from the outside.

🚫 What the rules actually prohibit

The specification — T/CAPT 026—2026, led by Xinhua and the Xinhua National Key Laboratory — sets an operating standard across the full chain: intake, content review, corpus access, optimization execution, distribution control, monitoring and traceability, risk handling.

Three practices are explicitly banned:

1. Corpus poisoning — polluting the data a model draws on.
2. Fabricated consensus — manufacturing the appearance that many independent sources agree.
3. Answer dominance — engineering artificial control over what an AI answer says.

And three requirements apply to every engagement: facts must be bound to evidence, sources must be tier-managed, and every operation must leave a traceable record.

Two regulators moved in parallel. China's cyberspace regulator launched a campaign to clean up AI application abuses that, for the first time, names GEO information poisoning and bulk source fabrication as priority targets. Separately, the China Advertising Association for Commerce began drafting a national GEO industry standard.

🔀 Two paths, visible in the contracts

The investigation found the market dividing along a clear line.

The legacy path still sells volume. One brand marketing lead told reporters that its GEO vendor charges over ¥200,000 (roughly US$28,000) a year — and still reports "media placement volume" as the core KPI. The deliverables: seventy or eighty press releases a month, near-identical in content, placed on obscure sites. The client's own description was that it is "no different from the Baidu news-source optimization of the past."

The compliance path sells verifiable authority. Institutions that built their GEO service on editorial process — reporter interviews, editor review, three-stage fact-checking — position themselves as a "custodian of local information credibility" rather than a distribution tool. Their monitoring scans how models cite the client's keywords daily, and when they find an error, such as an expired local policy being miscited, they push a corrected official source the same day.

The first model has a delivery unit that is easy to count and easy to sell. It also has no answer for the question that now matters: was the content actually cited, by whom, and can you prove where it came from?

📉 What happens when it goes wrong

The investigation documented a case that should end the debate.

A consumer brand hired an agency to push its ranking. The agency generated hundreds of fake user reviews with AI and fabricated quotes attributed to industry experts, aiming to manufacture the appearance of consensus. The short-term effect was real — the brand's presence in AI answers improved.

Then a major model platform updated its anti-manipulation system. The brand's citation weight was cut sharply, and according to the report, its appeal has still not been fully restored.

Two lessons follow. First, the penalty is not gradual. Compliance failures in generative engines do not degrade your ranking over time; they can remove your brand from the answer set at once. Second, the appeal path is opaque and slow — and for a company without a mainland entity, slower still.

🚩 The four warning signs in a GEO contract

This is where the investigation is most directly useful. If a vendor's proposal carries any of these four signals, the compliance risk is being transferred to you:

1. The KPI is placement volume. If success is measured in articles published rather than citations earned, you are buying the old product with a new label.
2. Guaranteed positioning. Contracts still promising "top-three brand mention in AI results" are offering something the specification explicitly bans. Absolute-effect guarantees are now a compliance red flag, not a selling point.
3. No audit trail. Vendors that will not provide an operations ledger, fact-verification records, or a change history are withholding exactly what the rules require them to keep.
4. Obscure placement sites. Content placed on unnamed, low-authority sites carries the least weight and the most risk — it is the same profile that triggers anti-manipulation systems.

Note the through-line: each signal shifts risk from the vendor to the client. That is not incidental. It is the business model.

🌏 Why overseas brands carry the most risk

The investigation's cases involved Chinese domestic brands, but the structural exposure is worse for a company headquartered outside China, for four reasons:

• Language. You cannot assess whether a Chinese press release is accurate, repetitive, or placed on a credible site without Chinese-language review capability.
• Vendor verification. There is no foreign-friendly registry of legitimate GEO providers. Distinguishing an editorial operation from a placement mill requires local context.
• Susceptibility to bad guarantees. "Guaranteed top three" is precisely the promise that reads as reassuring to a team that cannot verify results independently — and precisely the one the rules prohibit.
• Appeal asymmetry. If your citations are cut, your domestic competitor can work the platform relationship and the appeal process in Chinese, on the ground. You cannot.

This connects directly to what we covered in our September 10 analysis of ERNIE's citation behavior: the properties that carry the citations — Baidu Baike, Baijiahao, Baidu Zhidao — are precisely the places where posted content is checked against facts and traceability. Cutting corners there is not a shortcut. It is where enforcement lands.

🧭 What compliant GEO actually requires

Stripped of the marketing, the compliant model has three components, and none of them is a shortcut:

1. Authoritative sourcing. Content must come from, or be verifiable against, sources the models already weight highly — and under the tier system, those tiers are now formally defined.
2. Local knowledge depth. The credible operators compete on depth of local context — policy history, statistical records, primary documents — not on the number of articles they can publish.
3. Continuous correction. Monitoring is not a monthly report. It is daily scanning plus a mechanism to overwrite errors with current official sources, fast.

None of this is achievable through a foreign entity working alone. It requires a Chinese business presence, Chinese-language editorial capability, access to authoritative sources, and a compliance process that produces records an audit would accept.

BPP works at that intersection: Chinese-entity access to Baidu Baike, Baijiahao and Baidu Zhidao without a Chinese license, cross-platform AI visibility across the six Chinese assistants, and compliance-reviewed Chinese content built to produce the evidence trail the current rules require.

The market just demonstrated what the alternative costs. One brand's citations were cut, and its appeal is still pending.