ChatGPT still cites 50,000 fabricated statistics seven months after experiment, dev.to post reports

ChatGPT is still citing a website packed with 50,000 fabricated statistics roughly seven months after the site went live, according to a write-up on dev.to. The experiment, run by Metehan Yesilyurt of Peec AI, reportedly cost about $30 in GPT-4.1 API calls, and its author uses the results to argue that AI citation systems reward visibility rather than accuracy.

The experiment

According to the dev.to post, written by Juan Auriti, Yesilyurt built a site called stateglobe.com and populated it with tens of thousands of AI-generated statistics. Google later removed the site from its search index. Even so, the post claims, ChatGPT continued to surface and cite the site's content seven months after it was created.

The fabricated figures did not stay contained within chatbot answers. The post reports that the fake statistics have been referenced in an academic paper, in investment research from Qatar's largest bank, and by an Indian economics publication. More than 330 domains now link to the site, including Shopify, DHL and Bluehost, according to the author. The experiment also reached the front page of Hacker News on October 20, drawing around 350 points and 200 comments. All of these figures come from a single account and have not been independently corroborated.

How "citation laundering" works

The dev.to post labels the mechanism "citation laundering": fabricate data, get cited once, appear legitimate, and then get cited again in a self-reinforcing loop. The core vulnerability, the author argues, is that large language models cannot tell a fabricated statistic from a genuine one when the format looks authoritative. A number, a percentage sign and a confident sentence are treated the same as a figure from a peer-reviewed source, because the model extracts and synthesises rather than verifies.

Two consequences follow, in the author's framing. First, being cited by ChatGPT is not evidence of accuracy — a brand or source appears in answers because the model has seen it enough times, not because it has been checked. Second, any tool that counts citations without a trust layer is measuring something that can be manufactured for the price of a domain and some API calls.

The gap in AI visibility tools

The post widens the critique to the commercial "AI visibility" sector, naming Otterly, Peec AI, Scrunch, Profound, AthenaHQ, LLM Pulse, Evertune and Prefer as examples. According to the author, every one of these tools tracks citations or mentions, and none of them checks whether a citation is accurate, flags fabricated statistics, or traces a cited figure back to a primary source.

The proposed remedy is a trust layer with three components: verifying that claims attributed to a brand in AI answers are actually correct, checking the provenance of cited data against primary sources such as government databases or peer-reviewed research, and flagging any statistic that cannot be traced to an original origin. The author concedes none of this is easy to build, but argues the experiment shows it is necessary. It is worth noting that Peec AI, Yesilyurt's employer, is itself one of the visibility tools named in the piece.

The ethics of the stunt

The dev.to post is candid that the experiment was ethically questionable. Yesilyurt deliberately injected fabricated data into the AI citation ecosystem to demonstrate a weakness, and that data has since reached a bank's investment research and an academic paper, where it could influence real decisions. The author's defence is that the vulnerability needed exposing: if a small budget and several months are enough to plant tens of thousands of fake statistics into AI answers and earn hundreds of inbound links, the ecosystem is not self-correcting but self-amplifying.

Why it matters

This story matters well beyond the SEO and visibility-tool crowd. If AI assistants treat well-formatted fabrications as equal to verified research, then every answer that cites a statistic is a potential vector for misinformation — and the people reading those answers have no easy way to tell the difference. The reported spread into banking research and academia shows the blast radius is not hypothetical: synthetic data can migrate into human-authored documents that others then treat as credible. For anyone building on or reporting from LLM output, the lesson is blunt. A citation is a visibility signal, not a truth signal, and until verification layers exist, statistics surfacing in AI answers deserve the same scepticism as an unsourced claim from a stranger.