1. News
  2. AI models & agents
  3. AI search · Fake data
AI search · Fake data2 min read

A site of 50,000 fake statistics keeps getting cited by ChatGPT

A marketer's website of 50,000 invented statistics still earns ChatGPT visits and links months later, a reminder to check any number AI hands you.

By the Dashline deskSource: Hacker News (100+ points)
A site of 50,000 fake statistics keeps getting cited by ChatGPT
Photo: Alphaaloy, CC BY-SA 4.0

A marketer built a website of 50,000 statistics that were all made up, and ChatGPT is still citing it seven months later. The site, stateglobe.com, cost about $50 to build with AI in one day. Google dropped it. ChatGPT did not.

What happened

The author, who works at the AI search firm Peec AI, used GPT-4.1 to generate the pages in a single day. They covered topics like SEO, social media and e-commerce across 200 countries. Every number was invented, and none were checked. The first pages cost about $30. Later, 15,000 list articles added about $20 more.

The numbers so far

  • Google Analytics shows more than 10,000 visits from ChatGPT since March.
  • Over 330 websites link to it, including Shopify, Bluehost and DHL, according to Ahrefs and Semrush.
  • In the last 30 days, bots made 600,000+ requests. About 72,000 came from ChatGPT's live page-fetching agent.
  • On October 7, it ranked second for visibility on tracked prompts, behind Statista.

The author also says the fake figures showed up in an academic paper and in a bank's investment research. That claim comes from the author and is not independently checked here.

Why ChatGPT kept coming back

Google's visibility for the site collapsed after two updates that followed launch. Indexed pages now get only 5 to 20 impressions. But ChatGPT works differently. It rewrites a question into several smaller searches, pulls candidate pages, ranks them, then opens a few to build its answer. A page can be invisible on Google and still pass every one of those steps.

The author notes that the 72,000 requests came from the agent that opens a page while answering a real person's question. They were not routine data collection for training.

Who it affects

Anyone who asks an AI tool for a statistic and pastes it into a presentation, a grant application, a report or a blog post. Once a made-up number is published, others can cite it, and the chain grows. The experiment shows that a site with no real data can still look like a trusted source to an AI tool.

What to do

  • Ask the AI for its source, then open the source yourself.
  • Trace each number back to the original study or agency that produced it.
  • Be wary of a statistic that appears only on sites you have never heard of.
  • If you cannot find the origin, leave the number out.

It is not yet known how often ChatGPT repeats these invented numbers to everyday users, or whether OpenAI plans changes. The test is also one person's experiment, run by someone who sells tools in this field.

The takeaway

Before you use a statistic from an AI answer, find the original source yourself, because a made-up number can look just as convincing as a real one.

More articles

You should really read these.

Next step

Let’s talk