50,000 invented statistics, cited by ChatGPT: trace the number first

A researcher published 50,000 statistics he had invented with an AI model, on a new website, for about $50. Google dropped the site, but ChatGPT kept citing it for about seven months, and more than
Summary
A researcher published 50,000 statistics he had invented with an AI model, on a new website, for about $50. Google dropped the site, but ChatGPT kept citing it for about seven months, and more than 330 websites now link to it. A number an AI gives you is a lead to check, not a source you can cite.
The situation
Every week someone in my office, or a student in the Boardroom, pastes a figure from an AI answer into a draft. It might be the average margin for a pharmacy, the growth rate of a sector, or the share of MSMEs paid late. The figure looks reasonable, and there is a link beside it. It goes into a project report, a CMA projection for a bank, or a viva answer.
This month's experiment shows how little that link proves.
What happened
Metehan Yesilyurt, a researcher at Peec AI (a company that sells AI-search visibility services), describes the test in a post dated 7 October 2026. I re-read it on 24 October.
- He generated 50,000 statistics with GPT-4.1 in a single day, and later added 15,000 list articles. The cost was about $30 for the statistics and $20 for the lists. He writes: "Every one of them was fake" and "I never checked a single one."
- He published them on a new domain, stateglobe.com. After two Google updates the site's visibility collapsed, and he writes that Google has deindexed the pages.
- ChatGPT kept using it. He reports more than 10,000 visits referred by ChatGPT since March, and about 72,000 requests from ChatGPT's browsing agent in the last 30 days alone. On the prompts he tracks, the site ranked second for visibility, behind only Statista.
- The numbers spread. More than 330 websites link to it. He says the citing sources include an academic paper, investment research by Qatar's largest financial group, and an economics publication in India, which he does not name.
He calls it "citation laundering": every citation made the fake numbers look more trustworthy. The site now carries a disclaimer, and he writes that you should never publish fake content.
These are his own figures, from his own analytics, and nobody has audited them. The method is plausible, and the conclusion does not depend on the exact counts.
Where people go wrong
"It had a citation." An AI answer with a link only tells you where the AI found the claim. In this case the link led to a page built to be found, with nobody behind the number.
"Google would have caught it." Google did, and dropped the site. The AI answer kept it anyway. If your check is "it came up in search", you are checking the wrong system.
"Other people cite it too." That is how laundering works. A bank's research note citing a number does not make it true if the note took it from the same site.
"It is only a background figure." In a CA's work no number is only background. An industry margin in a project report shapes the loan. A growth rate in a valuation note shapes the price. If you sign the document, you own the number.
What to do
- Trace every number to the body that produces it, such as RBI, MoSPI, CBDT, GSTN, a ministry, or the company's filed accounts, before it goes into anything you sign.
- Cite that source and the date you read it, not the AI answer or the website that repeated it.
- Ask the assistant for the issuing body and the document, then open that document yourself.
- Use the client's own figures in projections, with the assumption written down, wherever an "industry average" would otherwise go.
- Tell your juniors and articled assistants the rule in one line: an AI number is a lead, not a source.
The tool
This is why the projection builder I made for bank loan files takes the business's figures only from what you type in, with each assumption printed beside the number. The bank-norm ratios it tests against are editable cells, and it has no "industry average" lookup. It is due an update before I link it, so there is no link here. It does not check whether your own assumptions are reasonable, and it does not know what the bank will accept.
Sources
- Metehan Yesilyurt, Peec AI — "Publish stats, get cited: the tactic that put my data all over the web" (7 Oct 2026) — https://peec.ai/blog/publish-stats-get-cited-the-tactic-that-put-my-data-all-over-the-web
Share line (for WhatsApp / LinkedIn — Karan pastes it once the article is live)
50,000 invented statistics got cited by ChatGPT for seven months, after Google dropped the site. An AI's number is a lead, not a source: trace it to RBI, MoSPI or CBDT. the related note