AI Generated 50,000 Fake Data Entries That Fooled ChatGPT for Seven Months, Cited in Academic and Investment Research Reports

In March 2026, researcher Metehan Yesilyurt spent about $50 to use GPT-4.1 to generate 50,000 fabricated statistics and built the website stateglobe.com, claiming to cover statistics on SEO, content marketing, and other fields across 200 countries.

What Happened

After the site went live, Google quickly reduced its visibility in two updates, and its indexed pages now receive only 5 to 20 impressions. ChatGPT continued to scrape data from the site; over the past 30 days, Cloudflare recorded more than 600,000 bot requests, including 72,000 from the ChatGPT-User agent. GA4 data shows that ChatGPT brought more than 10,000 sessions. The site currently has more than 330 backlinks, from domains including Shopify, DHL, and Bluehost. The fabricated data has entered an academic paper and a bank's investment research report.

Mechanism Breakdown

Experiments show that when generating answers, LLMs tend to grab specific numbers rather than vague statements, because numbers are more easily treated as authoritative evidence and repeated directly. The retrieval process is constrained by inference time and compute costs, and the model did not check the authenticity of sources. Google quickly reduced the weight of low-quality sites through algorithm updates, while ChatGPT's real-time retrieval mechanism did not simultaneously filter out content that had already been taken down, causing the same site to persist for markedly different lengths of time across systems.

Industry Impact

The citation of fabricated data in academic papers and financial institution reports shows that AI-generated content is entering formal decision-making chains through multi-level dissemination. If content producers rely on LLMs for research, they may amplify misinformation without knowing it. The different pace at which search engines and large model platforms assess data credibility increases the cost for downstream users to distinguish accurate information from false information.

Strategic Assessment

(The following is analysis, not fact.) From existing cases, relying solely on a model's own reasoning makes it difficult to solve the problem of confirming source reliability; in the future, external checking mechanisms or stricter retrieval filtering rules may be needed. When using AI tools, content producers need to establish additional manual review processes to reduce the risk of being contaminated by false data.