AI Large Models, Frequently Subject to "Poisoning"

Recently, the 2026 CCTV 3.15 Evening Gala revealed industry chaos caused by the abuse of Generative Engine Optimization (GEO) technology. Some service providers are mass-producing false information and systematically “feeding” AI large models to manipulate AI responses, even fabricating products to become AI-recommended “top picks.”

GEO is a marketing activity based on AI responses, where advertising companies produce and adjust content for brand clients to make it easier for AI chat software to capture and include in its answers, aiming to increase brand exposure or alter AI responses. As AI technology matures, the GEO sector has become increasingly popular.

According to reports, service providers engaged in GEO-related services claim that users only need to pay a fee to have their products listed in major mainstream AI large models, making their products appear as “standard answers” provided by AI models.

In an investigation, industry insiders purchased the “Liqing GEO Optimization System,” fabricated information about a smart wristband, generated and published over ten promotional articles, and then inquired about wristband recommendations on AI model platforms. Two major models ranked this fictitious product highly.

Li, the operator of the Liqing GEO system, said that the main reason GEO services are so popular is that they can feed or poison the AI large models to achieve clients’ commercial goals. To do GEO, the key step is to “publish articles” on various internet accounts. “For example, with mobile phone brands, there are five to ten positions. How do you handle so many phones? It could cost hundreds of millions in advertising annually, and spending a few million to inject some false information is acceptable.”

Several GEO company leaders told reporters that some content cited by AI is fabricated and requires users to carefully discern. Among these, rankings are a major source of false information, often created by companies to boost exposure after GEO efforts. There are many lists online ranking products or manufacturers, but these lists lack accurate factual basis and are likely fabricated, yet AI models often scrape these pages.

Previously, First Financial reported in an investigation titled “AI Flooded with Spam Marketing Info: Rankings Are a Major Issue, Changing AI Responses for a Few Thousand Yuan,” that optimizing a keyword costs about 3,600 yuan per quarter, and a year’s optimization costs around 10,800 yuan. Typically, a keyword is associated with over 500 related words, with platforms like DeepSeek as targets. Service packages often include bonuses like Doubao, Yuanbao, Tongyi, Kimi, etc. After signing a contract, results can be achieved within seven working days, with the brand name appearing in the model’s recommended sequence.

The development of the industry chain has driven the performance of the intelligent optimization sector in the secondary market. According to iiMedia Research, since late 2025, GEO concept stocks have remained “hot.” Lio股份 (002131.SZ) achieved four consecutive limit-ups in seven days, BlueFocus (300058.SZ) doubled its stock price in a week, Zhejiang Wenlian (600986.SH) and Yidian Tianxia (301171.SZ) hit daily涨停, becoming one of the hottest sectors since the start of the year. It is expected that by 2026, the industry will experience rapid growth, with a year-on-year increase of 159.3%, and market size reaching $29.17 billion.

However, the industry is also increasingly aware of the proliferation of low-quality online content. Mao Century General Manager Luo Xiaojun told reporters that there are cases of over-packaged content, uniform scripts, and large-scale distribution of marketing narratives replacing factual descriptions, making authentic content too weak in the information ecosystem. From the source of this spam information, aside from large companies hoping for stable AI citation of genuine data, some small and medium-sized companies unwilling to pay high marketing fees will seek GEO services, fabricating content to pollute the AI’s information sources and “fight” against the model’s rules.

Chen Jingjing, General Manager of the Yuwan Yinlan Integrated Marketing Light Consulting Division, said that GEO should return to the essence of brand marketing and create trustworthy content assets. “As AI continues to evolve and iterate, it will become smarter at discerning content quality. The public will also be further educated to improve their ability to identify the quality and source credibility of corporate GEO content.”

iiMedia Research analysis states that as a young sector, GEO still faces many challenges and has not yet reached a mature stage of development. Currently, industry standardization is low, effectiveness measurement standards are not unified, and service quality and results mainly depend on the platform and service providers’ capabilities, leading to some uncertainty in industry development. Whether GEO’s overall service model can sustain long-term value remains to be seen over time.

Reporter Zheng Xutong also contributed to this article.