A researcher created a website with 50,000 entirely fabricated statistics to test how large language models retrieve and cite information. The experiment, launched in March 2026, generated fake data on SEO, content marketing, social media, e-commerce, and web technology across 200 countries using GPT-4.1, costing approximately $50 to set up.
While Google quickly deindexed the site following algorithm updates, ChatGPT has continued retrieving pages from the domain regularly. According to Cloudflare data, the site received over 600,000 bot requests in the last 30 days alone, with approximately 72,000 coming from ChatGPT-User—the agent OpenAI uses to fetch pages in real time while generating answers. Over seven months, the site accumulated more than 10,000 sessions from ChatGPT and earned links from 330+ referring domains, including major brands like Shopify, Bluehost, and DHL.
The fabricated statistics have been cited in real-world contexts. According to the source, the fake data appeared in an academic paper, investment research from Qatar’s largest financial group, citations by an economics publication in India, and a research report from a Canadian influencer community. The researcher notes these represent only citations that linked back to the site; many more likely used the data without attribution.
The experiment highlights a critical distinction: retrieval by AI systems and ranking in Google search operate through different mechanisms. While Google visibility collapsed, the site maintained high retrieval marks in LLM systems. ChatGPT’s search process involves rewriting prompts into sub-queries, pulling candidates from multiple sources, and reranking results—a pipeline where a page invisible to Google can still perform well.
The researcher describes a concerning pattern termed “citation laundering.” A fake statistic on a new domain initially carries no authority, but once cited by academic papers or financial institutions, subsequent discoveries of that data show it with established sources. As cited pages enter web crawls and eventually training data for future models, fabricated information risks becoming embedded in AI training datasets.
The experiment was conducted as research to examine how LLMs ground answers in source material, not as a playbook for creating fake content. The researcher emphasizes that source verification should become essential practice.
Key facts
- A researcher created 50,000 fake statistics on a new domain to test AI retrieval patterns
- Google deindexed the site, but ChatGPT retrieved it 72,000 times in 30 days over seven months
- The fabricated data was cited in academic papers, bank investment research, and publications
- 330+ domains linked to the site, including major brands like Shopify and DHL
- AI retrieval and Google ranking operate through fundamentally different mechanisms