Researcher Metehan Yesilyurt of Peec AI conducted an experiment where he launched a website containing 50,000 entirely fabricated statistics across various industries. The site was generated using GPT-4.1 for approximately $50. While Google quickly deindexed the site following algorithm updates, ChatGPT continued to retrieve and cite its contents. Within seven months, these fake numbers appeared in academic papers, economic publications, and even investment research from major financial groups.
The image represents the primary source of the experiment conducted by Peec AI regarding the propagation of fake statistics.
The mechanism behind this phenomenon lies in how search-grounded LLMs perform information retrieval. When a user asks for a report or blog post, models often break the prompt into "fanout queries"—sub-queries specifically looking for keywords like "statistics," "data," or "%." Because a specific number provides more concrete grounding than a vague qualitative statement, the model’s reranking system identifies pages with high numeric density as highly relevant. Crucially, current LLM architectures operate on tight latency and cost budgets; they are designed to fetch and summarize content rather than verify every claim against independent primary sources before outputting a response.
This creates a risk of "citation laundering," where a single unverified statistic gains authority by being cited once by a reputable entity, which then becomes a source for subsequent users and models. This cycle can eventually lead to those figures being ingested into future training datasets. To mitigate such risks while still benefiting from AI visibility (GEO), Yesilyurt suggests that publishers should focus on providing clear, one-claim-per-sentence data points, matching headings to common fanout queries, and ensuring technical accessibility for bots through server-rendered HTML rather than hiding data behind JavaScript.