A site of 50,000 made-up stats is still being cited by ChatGPT
Metehan Yesilyurt, a GEO researcher now at Peec AI, built stateglobe.com in March 2026 with 50,000 statistics pages generated by GPT-4.1 in a single day. He never checked any of the numbers, and says the whole test cost about $50, including 15,000 listicles added later. Google soon dropped the site from its results, but he reports more than 10,000 ChatGPT sessions since launch, about 72,000 page fetches by ChatGPT-User in the last 30 days, and links from more than 330 sites according to Ahrefs and Semrush. He says the fake numbers ended up in an academic paper and in a bank's investment research. The site now carries a disclaimer, and he presents the work as research, not a playbook.
On Hacker News (315 points, 31 top-level comments) there is no consensus on what it shows. Some called it alarming and worth explaining to the public. Others said it is an old problem moving faster, citing Wikipedia's term "citogenesis" and low-grade outlets that exist to be cited. Two objected to the method itself, saying publishing fake data is simply lying. The author joined the thread to say he spent $50 in total.
Sorted take: a number with a link is not a checked number. Before a statistic from an AI draft goes into a deck or a client note, open the source and see where that source got it.
Why it matters: If your team drafts reports, pitches or blog posts with ChatGPT, make tracing every statistic to its original a rule. In this test a citation alone proved nothing.
Peec AI ↗Discussion on Hacker News ↗AI-assisted, reviewed by Sorted AI
▲ 0
0 comments