By Da Cheung
March 15 is world consumer rights day and in China that involves exposing business malpractice. This year, in its annual 3.15 TV consumer gala, China Central Television (CCTV) uncovered a thriving black market dedicated to “poisoning” large language models (LLMs). The investigation revealed how sophisticated Generative Engine Optimization (GEO) tactics are being used to manipulate AI chatbots into displaying fabricated content as objective truth.
To demonstrate the vulnerability of these systems, a CCTV reporter used an automated software system to invent a fictional product: the Apollo-9 smart band. The reporter generated and published a dozen promotional articles boasting exaggerated features such as “quantum entanglement sensors” and “bloodless glucose testing”. Within hours, this fictional device was being recommended by major domestic AI platforms, including DeepSeek and Doubao. The cost of this entire manipulation process was just 39.9 yuan ($5.50).
The software used, the Liqing GEO System, is openly available on e-commerce platforms. With a single “create article” command, the system automates drafting, image pairing, and distribution across a network of prepared social media accounts.
The industrialization of deceptive GEO
GEO is an emerging digital marketing framework increasingly used by companies to ensure AI models cite and prioritize specific content. It is gradually emerging as important as the more traditional search engine optimization (SEO) and while it can be used for legitimate marketing purposes, such as product updates, it is increasingly being weaponized to engineer a false consensus. By flooding the internet with fabricated articles and synthetic user reviews, companies can trick an AI’s cross-referencing logic into perceiving repetitive mentions as credible evidence and can push a product to the top of a generated answer.
According to the Chinese financial news investigative outlet Red Star Capital, industry insiders such as Hu Mingxuan are reporting a noticeable migration from traditional SEO marketing to GEO. Service providers charge different rates depending on keywords and platforms. Some offer a guaranteed top three placement in AI answers for an annual fee of 6,600 yuan. Others offer annual packages ranging from 2,980 yuan to 16,980 yuan. The premium package promises to churn out over 23,000 articles a year, relentlessly feeding content to manipulate search results for clients across industries including medical care, education, and home renovation.
One GEO service provider executive noted that its strategy is to build a chain of evidence to saturate digital ecosystem to the extent that the AI believes the information is both true and useful. Users seeking objective answers are instead unknowingly fed marketing or even deceptive material. Chinese market regulators have already issued guidelines for 2026 explicitly identifying AI-generated advertising as a key focus for oversight.
The next frontier: from search to answers
GEO is rapidly becoming the dominant technique in the marketing industry, representing a fundamental shift in how ordinary consumers acquire information.
In the past, search engine companies like Google made extensive efforts to combat manipulative SEO tactics, deploying algorithm updates to ensure more authentic and effective search rankings. However, today, major LLM developers are primarily preoccupied with upgrading their models, acquiring training data, and expanding market share. That’s left a vacuum where manipulative GEO can thrive and he profound impact this will have on users has yet to be addressed.
The looming threat of data pollution
Beyond intentional data poisoning, LLMs face a broader crisis: the exhaustion of easily accessible, high-quality human-written information.
As premium data becomes scarce, many models have begun training on content generated by other AI systems. Tech news outlet TMTPost reports that international industry leaders such as OpenAI and Anthropic have warned that the depletion of high-quality training data is a looming threat to the industry.
If fraudulent marketing content acts as an acute poison for an LLM, the massive influx of low-quality, machine-generated content is chronic pollution. Researchers call this model collapse, where AI models continuously learn from polluted data until they drift away from reality. As the industry pivots from competing merely on computing power to establishing and managing verifiable knowledge systems, the battle for clean data and authentic answers is just beginning.
Sources
