Fake-stats site pulled 72K ChatGPT-User requests in 30 days
Metehan Yeşilyurt's StateGlobe experiment shows LLMs reach for structured numbers when their training data falls short.
In March, practitioner Metehan Yeşilyurt built StateGlobe, a site of more than 50,000 invented marketing statistics by country. Chris Long summarised the results on LinkedIn on 9 October. The site exists to test whether LLMs will fetch original-looking numbers to fill gaps.
Google killed the site within about a month, Long reports. AI crawlers did the opposite. Cloudflare logs over seven months show 2.4M requests and 350K unique visitors. In the last 30 days, 670K requests hit the site, with 72K from ChatGPT-User, the agent tied to user prompts.
Semrush data cited by Long shows more than 300 referring domains picked up the fake numbers, including Shopify, DHL and Bluehost. One statistic landed in an academic paper. Another made it into an investment paper.
What the pattern actually shows
LLMs reach outside their training data when a prompt asks for a specific number they do not hold. Structured, fetchable pages get pulled. Writers using those models then paste the output into real articles, and the citations compound.
We think the honest reading is narrower than "statistics win citations." The site worked because models had a gap and StateGlobe filled it with clean, parseable answers to very specific questions. Country-level A/B testing rates are not covered well anywhere else.
Before you act on this
Check your own server logs for ChatGPT-User, PerplexityBot and Google-Extended before changing anything. The volume varies wildly by topic. A site serving queries models already answer well will see very little, no matter how much data it publishes.
Where a gap exists, the useful response is to publish primary data your readers would want anyway, with clear methodology and dates. That is a slower route than fabricating 50,000 rows, and it is the one that holds up when a human checks the source.