GEO in China: 90 billion yuan for AI visibility, some of it for poisoning models

A sector report from the portal ITHome describes one of the fastest-growing categories in Chinese marketing. GEO, short for generative engine optimization, is the work of getting a brand's content cited in language model answers. In 2026 the market is set to pass 90 billion yuan, with annual growth of about 160 percent.

User behavior explains the interest. China's CNNIC counted 560 million users of generative AI. A consumer survey indicates that more than 80 percent of customers check product information in a model before buying. Gartner forecasts that in 2026 half of the queries traditionally sent to search engines will be handled by AI, and that traffic from classic SEO will fall by roughly 30 percent a year. For marketers the arithmetic is simple: the budget follows the attention.

Who runs China's assistants

QuestMobile data from June 2026 shows what visibility is at stake. The leader is Doubao, ByteDance's assistant, with 382.3 million monthly active users. Then come Qwen with 167.18 million, DeepSeek with 129.82 million and Yuanbao, Tencent's assistant, with 49.84 million. Each of these apps is a separate world that has to be persuaded to cite a given source. Each also applies its own criteria for judging credibility.

Chinese practitioners talk about nine dimensions of evaluation. They say models have to be fed content that looks like reference material: with numbers, institution names, dates and citations. From there it is a short step to abuse.

A fake report, real citability

The portal 36Kr described how models are deliberately poisoned. A marketing agency prepared a fake report attributed to IDC about a certain accessories brand. It sent the report around portals that publish industry material and linked to it from dozens of sites. When a user asked the assistant for a ranking in that category, the model quoted the fabricated data as findings of a well-known research institute. The names of research institutions work like a credibility token. Repetition alone is enough for a model to treat the content as worth citing.

A second effect described is the mutual contamination of media and models. An AI search engine reads an article in the press, and a journalist quotes the assistant's answer. After a few rounds it is impossible to establish which source was the original one. Newsrooms lose control over the flow of information not because someone hacked them, but because the system took a fake for a fact.

Three labs, three answers

OpenAI, Google and Anthropic independently tightened their rules on citing sources in answers over the past year. All three try to favor outlets with a real editorial history and to prefer primary citation. That is a partial fix. As long as a ranking cannot be verified externally, it will be worth as much as the cheapest manipulation costs. The Chinese market shows that with the right motivation, that cost is very low.