Tolz.me
AIResearchFact Checking

50,000 Fake Statistics and ChatGPT Kept Citing Them: How to Check a Number Before You Use It

A site published 50,000 invented statistics in a day, and ChatGPT kept citing them. Here is how to check any number before you use it.

Tolz.me team7 min read
Illustration of a magnifying glass examining bar and line charts marked with question marks

In March 2026, Metehan Yesilyurt started building a site called stateglobe.com to test what happens when AI tools go looking for statistics. Every number on it was invented.

Yesilyurt wrote it up on October 7, 2026, on the Peec AI blog, under the title Publish stats, get cited. The post is worth reading for the numbers in it, and worth reading even if you never plan anything like this, because it shows exactly how a false claim travels now.

How it was built

  • 50,000 statistics pages generated in one day with GPT-4.1.
  • Every number on those pages invented and unchecked.
  • A cost of about $30 for the pages, plus roughly $20 more for 15,000 listicles.

The idea behind it: as more articles get written with LLMs, those models go looking for numbers to fill gaps in what they know, and statistics pages are exactly what they search for.

What happened next

On the search side, the plan failed. Google's visibility collapsed after two updates, and the pages were deindexed.

On the assistant side, it kept working. In the 30 days before the post, stateglobe.com received more than 600,000 bot requests. About 72,000 of them came from ChatGPT-User, which is the user agent ChatGPT uses when it fetches pages in real time to answer a question.

Analytics on the site showed more than 10,000 sessions from ChatGPT since March. Ahrefs and Semrush showed more than 330 referring domains, including Shopify, DHL, and Bluehost.

The invented numbers reached real publications

Yesilyurt lists where the made-up numbers turned up: an academic paper, investment research from Qatar's largest financial group, an Indian economics publication, and a Canadian influencer community report. He notes these are only the citations that linked back to the site.

The site now carries a disclaimer.

Why one invented number travels so far

A statistic page is the easiest kind of source for an assistant to use. It offers one number, one topic, and no argument to weigh. When a model needs a figure for a question, a page that contains exactly one figure fits perfectly, and nothing on the page contradicts it.

Then the number leaves the page. It lands in an answer, the answer gets pasted into a document, and the document cites the page. At that point the claim has a chain of custody that looks ordinary, and the original page is still sitting there for anyone who follows the link.

Retrieval is not ranking

The part worth keeping from the write-up is his split between retrieval and ranking. The site lost its search visibility completely, and that barely dented the citation flow. Assistants pulled the pages because the pages answered the question, not because those pages ranked for it. Google rankings and being retrievable by an LLM, he argues, are two different problems.

He also notes that LLM pipelines do not fact-check numbers. A number gets retrieved, quoted, and cited, and the check that would catch a false one happens somewhere else or not at all. He calls the result citation laundering.

And he is plain about the lesson for himself: you should never publish fake content.

How to check a statistic before you use it

  1. Find the primary source. Search the exact number in quotes. If it only shows up on aggregator sites and news summaries, keep digging for the study, dataset, or filing underneath.
  2. Read who published it. A number from a press release, a vendor survey, or a marketing page needs different handling than one from a peer-reviewed paper or a government dataset.
  3. Check the date. Statistic pages get republished for years. A figure from 2019 sitting in 2026 prose may describe a market that no longer exists.
  4. Ask what was measured and who was asked. Sample size, population, and collection method decide whether the number means what your sentence says it means.
  5. Check that the source says what you are claiming. Percentage of what, out of how many, compared with which baseline.
  6. Trace one step further. If you are citing a summary, open the summary's source. If that source points back to an aggregator, the chain ends there.

One practical note on step one. Text copied from a web page or a chat window can carry invisible characters, which can make a quoted search miss. If a quoted search comes back empty, run the text through the invisible character remover to see what the string really contains.

What to check on the page itself

  • No named author and no publication date.
  • No methodology section, no sample description, no definition of what was measured.
  • One number per page, with no other figure to compare it against.
  • A figure with no unit, no population, and no time period attached.
  • Writing that reads like a summary of a summary and cites no source of its own.

The same checks, applied to an AI answer

  1. Open the citation and look for the number on the page. If it is not there, stop and treat the figure as unverified.
  2. Look for the primary source behind the citation, not only the citation itself.
  3. Check that the number in the answer matches the number in the source, including units and time period.
  4. Find a second independent source before you publish, ideally one with a different method.

ChatGPT's own behaviour is part of the story here. When it fetches a page in real time to answer a question, the answer can rest on a single page it read seconds earlier, with no second source to check it against.

Warning signs in an AI answer

  • A precise number with no named source.
  • A link that turns out not to mention the number when you open it.
  • A figure that is unusually round, or unusually precise for the topic.
  • The same number appearing across several sources that all trace back to one origin.

Keeping a source log

Most false statistics reach a finished piece because nobody owned the verification step. A cheap fix is a short record kept next to the draft: the number as written, the source, the date you checked it, and the sentence in the source that supports it.

That last field is the one people skip, and it is the one that saves you. If you cannot point at the line that supports a number in your own draft, you do not have a source yet.

A rule for the draft

One rule covers most cases: do not paste a number into a draft until you have opened the source. Keep candidate figures in a separate list with the link attached, and promote a number into the text only after you have read the part of the source that supports it. It slows the first minute and saves the last hour of edits.

The same rule applies when a colleague sends you a number in a message. Ask for the source in that same thread, not three days later when the figure has already reached three decks.

What to do when the number turns out to be fake

Say so plainly and correct it. If you published it, fix the page and note the change. If somebody else published it, tell the author. A correction costs a paragraph. A statistic that spreads because you stayed quiet costs more.

If the number came from an assistant, the correction also belongs in the thread where you first pasted it. That is where the copies started.

How Tolz.me handles numbers

We do not publish invented statistics, and we do not republish a number whose source we have not opened. When a guide here needs a figure, it comes from a source we link, with the publisher named. The blog collects the rest of our guides.

Sources

Tools used in this guide

Open a tool, make the change, and copy the result. No account needed.