Publish fake stats, get cited in LLMs

A researcher built a website in March 2026 containing 50,000 statistics pages generated by GPT-4.1 - every number invented - and added 15,000 listicles, all for roughly $50. Google rapidly dropped the site from visibility, but LLM-powered search kept retrieving it: over seven months Cloudflare logged hundreds of thousands of bot requests (about 72,000 from ChatGPT-User in a recent 30-day window) and GA4 showed 10,000+ ChatGPT-driven sessions. The fabricated stats acquired 330+ referring domains, including major brands, and were cited in an academic paper, bank research, and other reports. The experiment demonstrates that LLM retrieval differs from traditional search ranking: models issue fanout subqueries for statistics, rerank candidates for relevance (not truth), open a few pages to ground answers, and rarely verify numbers due to latency and cost constraints. Those dynamics let fabricated figures gain apparent authority through citation laundering and risk seeding future training data.

A follow-up test using verified, properly sourced statistics produced similar link-generation results, confirming that well-constructed stats pages attract LLM citations ethically. Key takeaways are that specific, year-stamped numerical claims serve as strong retrieval signals, models prioritize freshness and quotable numbers, and content creators should make each statistic a standalone, citable passage while instituting rigorous source verification to prevent misinformation amplification.