Fake Stats Fool ChatGPT — And It Cites Them Everywhere
A researcher created a site with 50,000 pages of entirely fabricated statistics. Seven months later, ChatGPT was pulling from it constantly — and real institutions were citing it as fact. The implications for AI reliability are serious.
The Statistic That Wasn’t There
Metehan Yesilyurt noticed something uncomfortable while watching people interact with AI tools. When faced with knowledge gaps, ChatGPT didn’t say it didn’t know. It went looking for statistics — any statistics — to fill the hole.
So he built a trap.
Yesilyurt, who works on AI search tool development, purchased the domain stateglobe.com and used GPT-4.1 to generate 50,000 pages of fabricated statistics covering SEO, content marketing, social media, and e-commerce across 200 countries. The numbers were invented in a single day. He never verified a single one.
The site promised authoritative global data. It contained none.
What Happened Next
At first, things looked normal. Traffic came in, Google indexed pages, and the site appeared to perform reasonably. Then Google rolled out search algorithm updates that crushed organic visibility — the site dropped to a few dozen impressions per day from human searchers.
Yesilyurt stepped away for seven months.
When he returned, he found something that should alarm anyone who trusts AI to retrieve factual information. Bot traffic had exploded. In the most recent 30-day window alone, stateglobe.com received over 600,000 access requests. Approximately 72,000 of those came directly from ChatGPT’s own agent — labeled “ChatGPT-User” — pulling data to answer user questions.
The site had accumulated links from more than 370 other websites by the time he published his findings. Those links included pages from Shopify and other established brands. Yesilyurt noted the number would likely have grown further if Google’s search performance hadn’t collapsed.
Using a search visibility tracking tool, he found stateglobe.com ranked second in data visibility among all statistics sources — ahead of Mordor Intelligence and Similarweb, two well-known market research firms.
Why This Matters
The conventional warning about AI hallucinations goes something like this: AI sometimes makes things up. That’s true but incomplete. What Yesilyurt’s experiment demonstrates is more troubling — AI is actively selecting, retrieving, and citing fabricated content as if it were legitimate source material.
This isn’t a case of ChatGPT inventing numbers from thin air. This is a case where a bot deliberately crawled a website full of false data, treated it as a credible source, and reproduced it in answers to users. Then other websites linked to it. Then actual institutions cited it.
The citation cascade is the critical mechanism here. Academic papers referenced stateglobe.com. Qatar’s largest financial group used its data in investment research. An Indian economics magazine published it. Each citation adds a layer of perceived authority, even though the original numbers were never verified by anyone.
Yesilyurt built a second site with real, sourced statistics and observed the same pattern — AI traffic increased identically. The mechanism doesn’t discriminate between truth and fabrication. It discriminates between accessibility and obscurity.
How AI Chooses Sources
The reason this worked lies in how AI search agents actually operate. Yesilyurt explained that ChatGPT doesn’t simply copy the top ten Google results when answering questions. It rewrites the user’s query into sub-queries, pulls candidates from search sources and caches, then re-ranks those candidates to construct an answer with cited sources.
This means the path from a fabricated statistic to a cited answer in ChatGPT involves multiple layers of autonomous decision-making. The AI isn’t being tricked by a single bad page — it’s being systematically steered toward low-quality sources through the very architecture of its retrieval process.
Google search rankings and AI searchability are not the same thing. A site can be invisible to human searchers and highly visible to AI agents. Stateglobe.com proved this directly.
The Verification Gap
Yesilyurt’s conclusion is straightforward and underappreciated. AI systems operate under strict time and cost constraints. They don’t verify numerical accuracy by design — it’s computationally expensive and slow. Source verification remains the responsibility of human researchers and writers, not the machine.
That’s not a flaw that can be patched with a software update. It’s a structural limitation of how these systems are built to function at scale.
The experiment also reveals something about the economics of attention. Fifty thousand pages of AI-generated text, published in one day on a single domain, beat established market research firms in AI visibility metrics. Content volume and domain structure matter more to AI retrieval systems than factual accuracy — at least currently.
What Comes Next
Stateglobe.com now carries a disclaimer stating the data is fictional. But the damage from unattributed citations has already occurred, and the site continues to receive tens of thousands of AI-driven requests monthly. Other researchers have replicated parts of this approach — publishing fabricated data to observe AI behavior — suggesting this isn’t a one-off curiosity but a replicable pattern.
The broader implication extends far beyond statistics. Any domain where fabricated data can be published on accessible websites — health claims, economic forecasts, scientific findings, business projections — faces the same vulnerability. AI agents will retrieve, cite, and amplify that content without independent verification.
Human journalists, researchers, and analysts who rely on these tools need to treat AI-sourced data the same way they’d treat an anonymous comment on a forum: worth investigating, not worth trusting. The onus has shifted back onto the human side of the workflow — and that shift was already happening, but this experiment shows exactly how easily it can be exploited.
Yesilyurt emphasized that publishing false content was never the goal. The site was built as an experiment. But experiments like this reveal structural weaknesses that bad actors can exploit far more deliberately than a researcher testing a hypothesis.