On March 15, 2026, China’s annual Consumer Rights Gala — our most-watched consumer protection broadcast — investigated a business called GEO (Generative Engine Optimization). GEO providers charge clients to plant promotional content across the internet so that AI chatbots pick it up and recommend their products as if giving objective advice.

So how do you ‘poison’ a LLM?

CCTV reporters contacted multiple GEO firms. One, led by a manager surnamed Wang, claimed 200+ clients across industries within its first year. Wang explained the method: they produce advertorial articles on behalf of clients and publish them across platforms where AI models crawl and index content. Because AI algorithms update weekly, Wang said, they must continuously publish fresh content — “feed it, massively feed it” — to maintain rankings.

Another operator, surnamed Cheng, was blunt about why it works: “People don’t realize it’s advertising. That’s why they trust what AI tells them.”

A third, surnamed Zhang, described the goal as building a fake “evidence chain” across multiple sources so the AI cross-references them and concludes the client’s product is genuinely superior.

Pricing ranged from 299 yuan ($40) budget packages to 6,600 yuan ($900) annual contracts promising top-three placement in AI responses.

The reporters verified the claim

To test the claim, an industry insider bought a GEO tool called the “Liqing GEO Optimization System” from an e-commerce platform for 39.9 yuan (~$5).

They invented a nonexistent product: the AstroTekk Apollo-9 smart wristband, with fabricated features including “quantum entanglement sensing” and “black hole-level battery life.”

The software auto-generated over ten articles — product introductions, expert reviews, industry rankings, user testimonials — all containing the fabricated claims, including made-up user feedback and fake scores ranking the product first in its category. The tool then auto-published articles to pre-registered social media accounts, handling titles, body text, and images without human intervention.

Two hours after publishing, the insider queried a major Chinese AI model: “How is the Apollo-9 smart wristband?” The model responded with a detailed recommendation, repeating the fabricated claims verbatim and citing the freshly planted article as its source.

Over three days, the insider published 11 articles total — 8 “expert reviews,” 2 “industry rankings,” 1 “user review” — all generated and distributed by the Liqing system. When they then asked AI models to recommend a smart health wristband, two major platforms featured the fictional Apollo-9 with prominent placement.

The Operator’s Own Words

Li, the operator behind the Liqing system, described the business to the undercover reporter in explicit terms:

“Everyone in GEO is poisoning. There’s too much poisoned information online, so online information just isn’t accurate anymore.”

When the reporter asked whether this was harmful, Li acknowledged it was, then added that merchants also use the same tools to smear competitors — planting negative content about rivals through the same pipeline.

Li also described a downstream ecosystem the GEO boom had created: platforms specializing in bulk article placement, publishing hundreds of articles per day at a few dozen yuan each, serving as the distribution layer for AI data poisoning.

Aftermath

The day after the broadcast, a Bilibili creator replicated the same attack — different fake product, same method. The AI models fell for it again.

When users now ask AI models about the Apollo-9, the models correctly identify it as the fabricated product from the 3·15 exposé. Not because the models learned to detect the deception — it is simply the post-broadcast media coverage and public discussion became the new dominant source material on the topic.

References

央视财经 (CCTV Finance). “3·15晚会丨AI大模型，遭’投毒’” [3·15 Gala: AI Large Models, “Poisoned”]. WeChat Official Account, March 15, 2026. https://mp.weixin.qq.com/s/H9N_I7aSo28mcFVKr3ytIA

酷玩实验室 (Coollabs). “315曝光’AI投毒’：只需要10篇软文，就能把AI忽悠瘸了？” [315 Exposes “AI Poisoning”: Just 10 Soft Articles Can Fool AI?]. WeChat Official Account, March 19, 2026. https://mp.weixin.qq.com/s/CYjIjOLLjetbNrJ4LFT7nA