AI GEO Scams Boom, Classical PR Returns

🔑 Enhanced Key Takeaways

• •Attackers compromise high-authority websites like government and university sites to host GEO/AEO-optimized spam PDFs and content, exploiting LLM summarization to recommend fake support numbers.[1]
• •A Moscow-based 'Pravda' disinformation network flooded the internet with pro-Kremlin articles by March 2025, infecting AI chatbots' training and RAG datasets, with models repeating narratives 33% of the time.[3]
• •Researchers demonstrated in late 2025 that just 250 crafted documents can introduce permanent bias in LLMs, regardless of model size, enabling efficient well-poisoning for IO campaigns.[3]

🛠️ Technical Deep Dive

• •GEO/AEO poisoning involves injecting structured scam data (e.g., phone numbers, Q&A snippets in JSON-LD or PDFs) into compromised sites and user platforms like YouTube/Yelp to boost retrieval and direct quoting by LLMs.[1]
• •PoisonGPT proof-of-concept modifies open-source models to embed undetectable falsehoods (e.g., 'Eiffel Tower in Rome'), passing standard tests while skewing outputs on targeted topics.[2]
• •NYU study (Jan 2025) used GPT-3.5 to generate 50,000 fake medical articles injected into the Pile dataset; 0.01% poisoning increased harmful advice by 11.2%, with 1M poisoned tokens (0.001%) raising it by 5%.[5]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline