A $50 site of invented statistics reached a bank's research. Checking the number is now the job

Peec AI published an account on 7 October by its GEO researcher, Metehan Yesilyurt, of an experiment he had run for seven months. In March 2026 he bought a new domain, stateglobe.com, and used GPT-4.1 to generate 50,000 statistics pages in a single day. Every number was invented. He never checked a single one. The pages cost him about $30, and 15,000 listicles added later took the total to about $50. After two Google updates, the site vanished from Google. ChatGPT kept using it.

How a guess becomes a source

Yesilyurt's own crawl data shows the scale. In the last 30 days, Cloudflare logged more than 600,000 bot requests to the site. Around 72,000 came from ChatGPT-User, the agent ChatGPT uses to open a page in real time while it answers someone. GA4 shows more than 10,000 sessions from ChatGPT since March. Ahrefs and Semrush count links from more than 330 other sites, including Shopify, DHL and Bluehost. On 7 October, Peec's own dashboard ranked stateglobe.com second for visibility on the prompts it tracked. Only Statista ranked higher.

Then the numbers left the site. An academic paper cited them. So did investment research by Qatar's largest financial group, an economics publication in India and a research report from a Canadian influencer community. Yesilyurt counts only the citations that linked back, so the true spread is wider.

He calls this citation laundering. On day one, a fake figure had no authority. Once a paper and a bank had cited it, the next reader saw a statistic with a respectable source behind it. Each citation made the number look more trustworthy, and none of them made it true. His bigger worry is that these pages end up in web crawls and then in training data, at which point a GPT-4.1 guess becomes something the next model "knows".

Remember who is telling the story. Peec AI sells visibility in AI search, and the post ends as a guide to building statistics pages that get cited. The evidence stands. Peec also profits when the mechanism is widely understood.

Why the machine will not catch it

Yesilyurt describes the mechanism plainly. When a model researches a topic, it breaks the prompt into sub-queries and adds words like "statistics", "data" or the current year. A specific number is the most useful thing it can find. A grounded answer works to a tight budget for time and cost, and fact-checking happens nowhere in the process. Rerankers score relevance. If a passage matches the query and looks authoritative, it gets used.

For operators, this is the useful part. AI does the routine work. You do the thinking. AI tools already handle drafting, searching and assembling a deck. The step the system skips by design is checking the number, and that step needs a person who can be held accountable for it. Whoever moved stateglobe figures into a paper and a bank's research never traced them back to an origin, because there was no origin to find. That gap is where judgement earns its keep.

Your Next Move

1. Trace every figure before it goes into anything with your name on it. Follow each statistic back to the organisation that measured it and to its method. If the trail ends at an aggregator that names no method, cut the number. A missing figure costs you less than a fake one.
2. Audit what your team's AI drafts cite. Take ten AI-assisted documents from last month and check the source behind every statistic. Write down how many fail. That number is your case for adding a verification step to the workflow, and it is a better case when you are the one who found it.
3. Make the checking visible. When reviews weigh AI use, verified output is impact a manager can see. Our guide on how to show impact from AI use before your next performance review sets out how to present it.