China’s market regulators are now treating “GEO” (generative engine optimisation) as an enforcement area. On 10 October the 21st Century Business Herald counted three administrative penalties against GEO vendors this year, and the latest one goes further than “you lied in an ad”. Shenzhen’s regulator says the vendor probed AI systems for their inclusion rules and then fed them invented rankings.
The fines are small. What matters is how the cases are framed, because any brand or agency doing AI-visibility work in China should pay attention to it.
The three cases
- Shenzhen (published 23 September). Shenzhen’s market regulator listed it among 14 online-trading enforcement cases, reported by Shenzhen News. The Bao’an bureau found that 分晰牛科技’s Bao’an branch ran a GEO service that used a system to detect AI models’ evaluation preferences and extract platforms’ scoring and inclusion rules. It then wrote leading prompts and published promotional copy that steered around negative reviews and contained false or misleading claims, distributing it across social platforms. The regulator called this helping others advertise falsely, ordered the company to stop and fined it ¥50,000. IT Home and the Herald cite Article 8 of the Interim Provisions on Online Anti-Unfair Competition and Article 9 of the Anti-Unfair Competition Law.
- Guangzhou Yuexiu (4 September). According to the Herald, a GEO firm mass-produced articles such as “2026 Guangzhou criminal defence firms TOP5” for two law firms. It invented 187 cases, a 9.6 satisfaction score and a 92.5 overall score. The fine was ¥120,000.
- Beijing Chaoyang (June). A GEO vendor claimed on its own website a 99.6 rating, 49% market share and a 99% renewal rate. The regulator said this interfered with the normal operation of AI search products and fined it ¥50,000.
Why the legal framing matters more than the fines
Two of the three cases are ordinary false advertising. The third is new: the regulator described the harm as damage to the AI search product itself. Cheng Nian, who heads AI law research at Zhejiang Kending (Beijing), told the Herald this falls under Article 13 of the Anti-Unfair Competition Law, its catch-all clause. A third party doesn’t need to touch a platform’s code. Manipulating the information an AI collects, and so interfering with the product, can be enough. In Cheng’s words, if a vendor “pollutes the information environment” of an AI platform and weakens its neutrality, the two may be in a competitive conflict, and that can amount to unfair competition.
There’s a precedent nearby. In May, the Herald notes, Hangzhou Intermediate Court ruled on appeal that a company selling an AI tool for mass-producing Xiaohongshu “seeding” posts had interfered with the platform’s normal operation, and awarded ¥100,000 in damages. Put that alongside the Beijing GEO case and Chinese platforms now have a legal argument for going after vendors that flood their answers with fabricated content.
How big is the problem?
The Herald’s own test in April found that more than 30% of over 1,000 reference links cited by domestic LLMs showed clear signs of GEO placement. Of the GEO content it identified, over 80% involved fabricated authority, data or reviews. An agency told the paper that small vendors often skip human review and batch-generate copy with AI, and can get it cited by a model “in as little as three days”. The same agency said client demand hasn’t cooled since state broadcaster CCTV’s 3·15 consumer-rights gala exposed “AI poisoning” this year, but clients now ask for operation logs and human sign-off.
That fits what we’ve seen from the platform side. An audit found Baidu’s AI overviews far more source-concentrated than Google’s, so a handful of fake “TOP5” pages can do a lot of damage. Tools such as Qwen Model Studio’s web search also put citations front and centre. The more concentrated and visible the sources are, the more it pays to fake one, and the more regulators will care.
What this means if you do AI visibility work in China
- Self-made rankings are the obvious target. All three cases involve invented scores or ratings, and one involved “TOP5”-style league tables. If your GEO plan relies on publishing ranking pages that put your client first, change the plan.
- Every number needs a source. Cheng’s checklist is simple: is the source real, is the expert real, can the data be verified, and is commercial promotion dressed up as independent review?
- Probing models is fine. Fabricating to exploit them isn’t. The Shenzhen case named the “GEO analysis system” in its findings, but the penalty came from what the vendor did with the results. Testing how models cite sources is research. Making up facts to fit those patterns is what got fined.
- Keep records. Logs, human review and approval trails are becoming the industry’s defence. Western agencies working for Chinese-market clients should build them in now.
My take: ¥50,000 won’t worry anyone, but the precedent should. Chinese regulators have accepted the argument that “manipulating the AI” harms the AI provider, not only consumers. That gives Baidu, Alibaba, DeepSeek and the rest a stronger basis for asking regulators to act. Clean, verifiable, original content was always the better long-term GEO strategy in China, and it’s now the lower legal risk too. More in China SEO.