Industry Practitioner Recounts: GEO Exposed in the March 15 Probe — How Did It Precisely "Dupe" AI?

CCTV 315 unmasked "AI poisoning" tactics

CCTV’s 315 probe has thrust a new marketing tactic into the spotlight: GEO (生成式AI搜索引擎优化). The show demonstrated how feeding mass, often fabricated, content to large language models can push a brand, product or service into AI-generated answers — even when the product does not exist. In one sting, a fictitious smart band called “Apollo‑9” was fed into a system called the "力擎GEO优化系统"; when reporters later queried an AI search tool, the non‑existent band was presented as “industry number one.” It has been reported that some GEO vendors advertise their services in hyperbolic terms — “manipulate AI,” “make AI obedient,” “brainwash AI” — claims that watchdogs and users have questioned.

What GEO does, and how it differs from SEO

GEO is not just old‑school SEO repackaged. Traditional SEO competes for click‑throughs on search engine result pages. GEO’s objective is different: make generative models treat a brand’s content as authoritative and emit it directly in conversational answers. Practitioners say the work involves three core moves — identify the real user questions, craft answers in an AI‑friendly structure (conclusion first, 3–5 supporting points, succinct evidential claims, a quotable summary), and place that structured content where models habitually crawl. Reportedly, vendors also perform technical fixes (robots.txt checks, Schema markup) and use PR‑style channel placement on high‑authority sites so the model “trusts” the source.

Industry debate, measurement headaches and a growing market

The industry is split. Some marketers insist GEO produces measurable commercial lift; others call it unstable, opaque and ripe for abuse. Pricing ranges from a few thousand to several hundred thousand yuan depending on scope and difficulty. Measurement is thorny: vendors often validate results by submitting bespoke prompts to ChatGPT or other AI search tools and counting brand mentions, but AI non‑determinism and personalization undermine such proofs. Overseas players are already productizing parts of the stack — US startup Scrunch AI, for example, has built infrastructure to convert content into model‑friendly structured data — showing GEO is becoming an international business, not just a China‑centric trick.

Why it matters, and what comes next

Why should Western readers care? Because the problem touches on universal issues: information provenance, model ingestion pipelines, and the incentives that reward visibility over truth. As generative AI becomes strategic technology amid trade tensions and export controls, regulators and platforms may tighten rules around content provenance and training data sources. For now the message from practitioners is blunt: GEO works when content is high quality, sector‑deep and well‑placed — but when those levers are abused, the result is not better answers but a new vector for misinformation. Who will police the feeds into the models — and how — is the urgent question left hanging.