GPT-4.1 spent $50 generating 50,000 fake statistics pages; ChatGPT kept citing them after Google removed the site
Published · updated · curated by AI Is Going Just Great
Source: news.lavx.hu ↗
Researchers and analysts can mistake the number for evidence once a bank note and an academic paper cite it.
A Peec AI researcher built a network of 65,000 fabricated statistics pages for about $50 using GPT-4.1, then watched ChatGPT continue to surface them long after Google deindexed most of the site. Over seven months, the fake network accumulated more than 10,000 ChatGPT referral sessions, links from 330-plus domains, and citations in an academic paper and investment research from Qatar's largest financial group. On October 7, 2026, StateGlobe ranked second behind Statista for a tracked set of prompts.
The gap Peec exposed is structural: Google's ranking system eventually removed the pages; ChatGPT's live-answer retrieval kept fetching them anyway. Rerankers score a passage for relevance, not accuracy, so a fabricated statistic formatted with a year in the title and URL looks fresher and more specific than a bare claim. Peec calls the downstream effect citation laundering — once a bank note or academic paper repeats a generated number, the institution's credibility travels with it, and crawlers can sweep those secondary citations into future training sets.