A marketer built a website with 50,000 invented statistics for about $50, generating them in a single day with GPT-4.1 and never checking a single figure. The site was launched on the new domain stateglobe.com and presented itself as providing SEO, content marketing, social media, e-commerce and web technology data from 200 countries.
Google deindexed the pages after two updates shortly after launch, and the site has been nearly invisible in Google results for months. ChatGPT, however, kept citing it. Over seven months the site has drawn more than 10,000 AI sessions and links from more than 330 referring domains, including Shopify, Bluehost and DHL. The made-up numbers have also surfaced in an academic paper and in a bank's investment research.
The experiment grew out of research into how large language models look for reliable data when they lack information. Because time-sensitive content is often written with LLMs, the models frequently search for statistics to fill gaps. The author argues that most generative engine optimization advice focuses only on getting a brand into an AI answer, while this test examines what happens after people reuse that output.
Cloudflare data showed more than 600,000 bot requests in the last 30 days, about 72,000 of them from ChatGPT-User, the agent ChatGPT uses to fetch pages. The results highlight a risk for anyone relying on AI search for facts: fabricated figures can spread widely once they are published, even when search engines no longer surface them.
Source: Hacker News · Summarized by HeadlinesBriefing