The experiment was elegantly simple. In March 2026, a GEO researcher bought a fresh domain, stateglobe.com, and used GPT-4.1 to generate 50,000 fabricated statistics pages across SEO, e-commerce, social media, and web technology topics for 200 countries. Total cost: roughly $50 including 15,000 listicles added later. No number was checked. The template was a basic AI web design shell. Google's immune system worked. Two updates hit shortly after launch, and the site's organic visibility collapsed to between 5 and 20 impressions. The site has been functionally invisible on Google for months. But ChatGPT never got the memo. In the last 30 days alone, the site received more than 600,000 bot requests, with 72,000 coming specifically from ChatGPT-User — the real-time retrieval agent that fetches pages while answering live queries, not a training crawler. GA4 shows 10,000+ ChatGPT sessions since launch. The downstream contamination is the real story. More than 330 referring domains now link to the site, including Shopify, DHL, and Bluehost. The fabricated numbers were cited in an academic paper, used in investment research by Qatar's largest financial group, referenced by an Indian economics publication, and sourced in a Canadian influencer research report. On October 7, 2026, stateglobe.com ranked second for visibility on tracked prompts in Peec AI's dashboard — behind only Statista, an actual data powerhouse. The mechanism is structural, not accidental. When an LLM is asked to write content, it breaks prompts into fanout sub-queries that append modifiers like "statistics," "data," or the current year. A page offering "X% of employees work remotely" beats a page saying "remote work is growing" because the number is specific, quotable, and fills an evidentiary gap in the model's output. LLMs are not evaluating truth — they are satisfying a format preference for numeric specificity. This creates what the researcher calls citation laundering. On day one, a fake number on a zero-authority domain has no credibility. Then an academic paper cites it. Then a bank. The next person or model encountering the stat now sees it backed by institutional sources. The number never became true, but each citation made it look more trustworthy. And after enough retrieval and citation, these pages enter web crawls that become training data — meaning a GPT-4.1 guess could become something the next model "knows" as fact. The key technical finding separates two problems most people conflate: Google ranking and LLM retrievability are independent systems. ChatGPT does not simply copy Google's top 10. It rewrites prompts into sub-queries, pulls candidates from search sources and its own cache, reranks them, and opens pages to ground answers. A page can be invisible on Google and still perform well across every LLM retrieval step. For the growing GEO industry, this means Google visibility is one signal, not the whole picture. The researcher replicated the tactic with real, properly cited statistics at peec.ai and observed the same link-acquisition pattern almost immediately — confirming the mechanism works regardless of data veracity. The uncomfortable conclusion is not that fake stats spread (they always have), but that the LLM retrieval pipeline has no meaningful truth-verification layer and actively selects for the format most likely to contain fabrications: precise-looking numbers on pages optimized for machine consumption rather than human scrutiny.
