Fake Statistics Gain Authority Through AI Citations

A researcher created a website containing 50,000 entirely fabricated statistics to study how false information propagates through AI systems. Built for roughly $50 using GPT-4.1 to generate fake data on marketing, SEO, and e-commerce across 200 countries, stateglobe.com was quickly deindexed by Google. ChatGPT, however, never stopped retrieving from it—generating 72,000 requests monthly to cite the false numbers as sources in real user conversations.

The experiment exposed a fundamental difference between traditional search ranking and language model information retrieval. Though invisible to Google, the site maintained strong performance in ChatGPT’s search pipeline, accumulating over 330 backlinks from major brands including Shopify, DHL, and Bluehost. More troublingly, the fabricated statistics migrated into genuine publications: academic papers, bank investment research, economics articles, and industry reports all cited the data as authoritative sources.

This dynamic creates what the author calls “citation laundering”—a feedback loop where invented information gains credibility through citations from reputable institutions, making each iteration appear more trustworthy. The deeper risk: as these recycled statistics enter training datasets for future AI models, misinformation could become embedded in the systems themselves. The findings reveal both how LLMs assess source reliability differently than search engines and the potential for AI to accelerate information cascades that solidify false claims.