He planted 50,000 fake statistics. Now developers are debating who checks AI's citations.

A website built around invented statistics is prompting a fresh argument about whether an AI citation is evidence of truth or merely evidence that a page was retrieved. Peec AI researcher Metehan Yesilyurt says he generated 50,000 statistics pages for StateGlobe with GPT-4.1, then watched their figures acquire links and apparent authority elsewhere on the web.

The experiment began in March, and his account was published October 7. The new hook is its October 20 Hacker News discussion, which showed 327 points and 192 comments when reviewed October 21. Those are observed discussion totals, not a growth measurement or evidence that the experiment just launched.

The author says Google’s rejection did not stop AI retrieval

Yesilyurt reports that StateGlobe’s Google visibility collapsed while ChatGPT continued fetching and citing pages. He says analytics showed more than 10,000 ChatGPT referral sessions since launch and backlink tools showed more than 330 referring domains. He also identifies examples of the invented figures appearing in an academic paper and investment research.

These are the experimenter’s claims and measurements. AITrending has not audited his analytics, traced every backlink or independently established how each downstream writer found the material. A bot request is not itself a user visit, a citation or proof that an answer repeated a number. The post distinguishes server requests from referral sessions, and those categories should remain separate.

The broader concern is what the author calls citation laundering: an unsupported number can look more credible once another publication cites it. His suggestion that such content could later enter training data is a possible downstream risk, not something this account demonstrates happened.

A real debate around a promotional research account

The English Hacker News replies debate incentives to cheat, the loss of source context in generated answers and whether similar failures already affect human readers. That discussion establishes current interest; it does not independently validate the experiment’s totals.

Peec sells AI search analytics, and the article promotes its own tracking tools alongside the research. Yesilyurt says StateGlobe now carries a disclaimer and explicitly advises against publishing fabricated content. The practical takeaway is to follow a statistic back to its underlying evidence. A linked page can be a genuine source of a claim while still being a poor source for believing it.