What is AI poisoning? The GEO grey market exposed on China's 3·15 consumer show
AI poisoning floods the web with fake articles so AI assistants cite them. What China's 2026 3·15 show exposed, how it differs from real GEO, and how to vet an agency.
AI poisoning doesn’t attack the model; it pollutes the web pages AI reads. Publish enough false or exaggerated articles and an assistant searching the web may treat them as references, putting a product that doesn’t exist, or isn’t good, into its recommendations. On 15 March 2026, China Central Television’s annual 3·15 consumer-rights show exposed a grey market doing exactly this under the GEO label.
What the 3·15 show exposed
According to the report republished by Guangming Online:
- Investigators bought software called “Lijing GEO Optimization System” on an e-commerce platform, used it to generate more than ten promotional articles for a fictional smart bracelet, and published them online.
- When they then asked AI for smart health bracelet recommendations, two large models listed the fictional bracelet near the top.
- A GEO agency head quoted in the report described the method as essentially writing advertorials and letting AI platforms ingest and crawl them.
The experiment shows that on questions with little information and weak competition, a handful of articles can sway AI answers.
How poisoning differs from legitimate GEO
| AI poisoning | Legitimate GEO | |
|---|---|---|
| Content | Invented products, inflated claims, fake reviews | Real product facts, specs, cases and limits |
| Publishing | Bulk advertorials posing as third parties | Your own site and bylined content; third-party mentions from real relationships |
| Measurement | A screenshot of “we’re listed” | A fixed question set run repeatedly, tracking mention rate and accuracy |
| Risk | Legal exposure for false advertising; downranking once detected; brand damage when users find out | Slower results; ongoing upkeep |
Generative engine optimization is meant to help AI understand a brand that really exists. Occupying answers with false content may work briefly, but it doesn’t survive scrutiny, and the cost lands on the brand when platforms change retrieval or users notice.
Vetting a GEO agency
- Where does the content come from? Is every product claim backed by your own material, and who approves it?
- Where is it published? Your site and named channels, or batches of anonymous articles and “independent reviews”?
- Be wary of guaranteed rankings or results in days. Agencies don’t control AI answers; promising a fixed position usually means shortcuts.
- Ask for raw evidence. The question, platform, date and full answer for every test, not just a ranking screenshot.
Why agency rankings disagree is covered in can you trust GEO agency rankings. To see how AI describes your brand today, start with a free AI visibility audit.
Frequently asked questions
What is AI poisoning in GEO?
Publishing large numbers of false or exaggerated articles so that AI assistants retrieve and cite them. It targets the web pages AI reads, not the model itself.
How do I tell if a GEO agency uses poisoning tactics?
Ask where each product claim comes from and who approves it, where content is published, and for raw test evidence. Be wary of guaranteed rankings or results within days; no agency controls AI answers.
Related
- Generative engine optimization (GEO) — Generative engine optimization (GEO) improves how AI answers describe, mention and recommend a brand, measured with fixed questions and retests.
- Mention rate — AI mention rate is the share of valid AI answers to a fixed question set that correctly mention your brand: correct mentions divided by valid answers.
- Can you trust AI product recommendations? How ChatGPT and other assistants pick brands — AI recommendations are answers built from training data and live search, not reviews or rankings. How they're made, how to check one, and what 'AI recommended' ads mean.
- Can you trust GEO agency rankings? How to read a 'top 10 GEO companies' list — Most GEO and AI SEO agency rankings are published by companies on the list, with no method, sample or date. Five questions to ask, and a better way to compare agencies.