GEO in China: 90 billion yuan for AI visibility, including model poisoning

A sector report from ITHome highlights one of the fastest-growing categories in Chinese marketing. GEO, or generative engine optimization, involves ensuring brand content is cited in language model responses. The market is projected to surpass 90 billion yuan in 2026, with annual growth of approximately 160 percent.

User behavior drives this interest. China's CNNIC recorded 560 million users of generative AI. A consumer survey indicates that over 80 percent of customers check product information in a model before purchasing. Gartner forecasts that in 2026, half of queries traditionally sent to search engines will be handled by AI. Traffic from classic SEO is expected to drop by roughly 30 percent a year. For marketers, the calculation is simple: budgets follow attention.

Who controls China's assistants

QuestMobile data from June 2026 shows what visibility is at stake. Doubao, ByteDance's assistant, leads with 382.3 million monthly active users. Qwen follows with 167.18 million, DeepSeek with 129.82 million, and Yuanbao, Tencent's assistant, with 49.84 million. Each app is a distinct ecosystem that must be persuaded to cite specific sources. Each applies its own criteria for judging credibility.

Chinese practitioners discuss nine dimensions of evaluation. They argue that models need content formatted like reference material: including numbers, institution names, dates, and citations. This creates a short path to abuse.

Fake reports, real citability

Portal 36Kr described the mechanism of deliberate model poisoning. A marketing agency created a fake report attributed to IDC regarding an accessories brand. They distributed it across industry portals and linked to it from dozens of sites. When users asked assistants for rankings in that category, the model cited the fabricated data as findings from a well-known research institute. Institution names act as credibility tokens. Repetition alone is sufficient for a model to deem content worth citing.

A second effect is the mutual contamination of media and models. An AI search engine reads a press article, while a journalist quotes the assistant's response. After several iterations, it becomes impossible to determine the original source. Newsrooms lose control over information flow not because of hacking, but because the system accepted a fake as fact.

Three labs, three responses

OpenAI, Google, and Anthropic independently tightened their source citation rules over the past year. All three aim to favor outlets with genuine editorial histories and prefer primary citations. This is a partial fix. As long as rankings cannot be externally verified, their value equals the cost of the cheapest manipulation. The Chinese market shows that with sufficient motivation, that cost is very low.