AI search found a website full of invented statistics
retrieval
Finding and opening a page for an answer, without proving its claims.
citation laundering
A false claim can look trustworthy after other documents cite it.
An October 7, 2026 Peec AI report describes an experiment with a troubling result. A site full of invented statistics remained visible to AI retrieval systems after Google visibility collapsed. The report is a company’s own account, so its measurements need independent checking.
What happened
In March 2026, Metehan Yesilyurt, a Peec AI researcher, created StateGlobe. GPT-4.1 generated 50,000 statistics pages in one day. They covered topics across 200 countries. Every number was invented. The first build cost about $30. Later, the researcher added 15,000 summary pages. The total cost reached about $50.
Google first indexed much of the site. Then its visibility fell sharply. Seven months later, Cloudflare logs showed more than 600,000 bot requests during one recent 30-day period. About 72,000 were labeled ChatGPT-User, according to the report. The article also says ChatGPT generated more than 10,000 sessions since March. More than 330 other domains linked to the site. The author says invented figures appeared in an academic paper and several research reports.
The background
Peec AI says AI search does not simply copy Google’s top ten results. It breaks a question into smaller searches. It gathers possible pages, reranks them, and opens some pages while building an answer. A page can rank poorly on Google while remaining retrievable by an AI system.
Specific numbers are useful in this process. They fill a missing detail and can be quoted easily. A number with a link can look like evidence, even when nobody checked the original research. The report calls the chain citation laundering. Later citations can make a false claim look increasingly serious.
Why it matters
This is not only a search-ranking problem. A page visit proves that a system found the page. It does not prove that the system accepted the claim. If a writer copies the number into a report, the citation may create the appearance of independent support. Another reader may then copy the same number again.
Peec AI says the experiment is research, not a playbook for fake content. It also says people should never publish fake information. That warning matters because the experiment shows how cheap and simple a misleading page can be.
What the report establishes
The report provides the site’s scale, cost, traffic records, and link counts. Peec AI also ran a second test with checked statistics and clear sources. It says other sites linked to that accurate page quickly. That suggests useful statistics can attract links. It does not prove that fake data will always spread.
The Hacker News discussion was listed with 294 points and 167 comments. Those figures measure community attention. They do not prove the experiment is correct.
What remains unknown
Access logs do not show whether ChatGPT used the invented numbers in user-facing answers. The report does not trace each path into the academic paper or research reports. People may have copied the figures directly. An AI system may have played no role.
The article also does not provide an independent audit of its traffic and link measurements. It is unclear whether other AI search systems would behave the same way. The experiment is important evidence of retrieval, not final proof of widespread belief.
What to watch next
AI search services may need to separate retrieval from verification. They could show the data owner, original study, date, and checking status. Writers and readers should follow links back to original research. They should check methods instead of trusting a precise number alone.
Peec AI says future systems may close this gap. The safest lesson is simple: a citation can show where a claim appeared. It cannot, by itself, show that the claim is true.
A website made up 50,000 statistics. AI search still found it.
📰 Full story: AI search found a website full of invented statistics
A page visit does not mean an AI knows the numbers are true.
AI search
A system that finds web pages and uses them to build an answer.
citation
A link or reference showing where a claim came from.
💡 The gist
- A company made 50,000 statistics that were all invented.
- Google hid the site from most search results.
- ChatGPT’s web reader kept opening its pages.
Peec AI studies how AI search finds information. Its researcher created StateGlobe in March 2026. GPT-4.1 made 50,000 pages in one day. The pages covered many topics and 200 countries. Every number was fake. The first build cost about $30. Later pages raised the total to about $50.
Google first indexed the site. Then its visibility fell sharply. Seven months later, the site’s logs showed more than 600,000 bot requests in one recent month. About 72,000 came from ChatGPT-User. This name refers to a tool that opens web pages while ChatGPT answers. Page visits do not prove that ChatGPT accepted the numbers.
Peec AI also reported more than 10,000 ChatGPT sessions since March. It counted more than 330 linking domains. The article says some invented figures appeared in an academic paper and research reports. It does not trace every step. People might have copied the figures directly. ChatGPT might have played no role.
Why could this happen? AI search often breaks a question into smaller searches. It looks for useful pages and specific numbers. A number feels easy to quote. A link can make the number look trustworthy. Later citations can make it look even stronger. The report calls this citation laundering.
Peec AI ran another test with checked statistics. Other sites linked to that accurate page quickly. This suggests that useful statistic pages can attract attention. It does not prove that fake data always spreads.
The Hacker News post had 294 points and 167 comments. That shows strong interest. It does not prove the experiment is correct. Readers should check the original study, method, date, and data owner. Peec AI says people should never publish fake content. It also says future AI systems may close this gap. Read the Peec AI report. See the Hacker News post.
💬 AI can repeat false numbers with a confident voice
The comments warn that an AI may not recognize a lie and may restate source-like web text as if it were reliable.
- If a web page looks trusted, an AI may repeat its data even when the data is made up.
- Checking another page may not help if that page copied the same false claim.
- One user reported that an AI said it had checked facts when it had not. That is a user report, not independent evidence.
- Proposed fixes include strict data rules, checking a claim’s origin, and using official first-party datasets such as BLS. The comments do not show that these fixes are foolproof.
- People lied online before AI, but AI can make the lies cheaper to copy and harder to trace. Some commenters want company accountability; others fear government control of information.
initial digest at 167 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
A computer found a website full of made-up numbers
📰 Full story: AI search found a website full of invented statistics
A page visit does not mean the computer knows the numbers are true.
ChatGPT
A computer helper that answers questions with words.
StateGlobe
The name of the website used in the experiment.
Peec AI studies how AI search finds things.
Its researcher made a website called StateGlobe. The site had 50,000 numbers. The numbers were all made up. The website cost about $50 to make.
Google stopped showing the site often. But ChatGPT, a computer helper for questions, kept opening pages. Bots asked for pages more than 600,000 times in one recent month. About 72,000 requests came from ChatGPT-User. That counts visits. It does not count beliefs.
The researcher said some made-up numbers appeared in research writing. We do not know who copied them. We do not know whether ChatGPT used them in answers.
The story became popular on Hacker News. It had 294 points and 167 comments. Popularity does not make a story true. We should look for the original study. We should ask who checked the numbers. Peec AI report. Hacker News discussion.
💬 AI may say made-up numbers as if they were true
People can put made-up numbers on the internet. An AI may not notice that they are made up.
- If a page looks trustworthy, the AI may use its false numbers in an answer.
- Looking at another page may not help if that page repeats the same lie. One user said an AI claimed it had checked the facts when it had not; that is only that user’s report.
- People suggest using official numbers, keeping track of sources, and giving AI checking rules. These ideas may help, but they are not perfect.
- People have always lied online. AI can make lies easier to spread, so some people want companies and rules to help, while others worry about too much government control.
initial digest at 167 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
💬 How fake data can become a credible-looking LLM answer
The central concern in the comments is not that every LLM deliberately lies, but that it may not know whether data is fake and can turn web text into a confident, source-like answer. Commenters propose safeguards, but disagree about their reliability, cost, incentives, and regulation.
initial digest at 167 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.