Fake stats site ignored by Google gets 10,000 ChatGPT sessions and real links
A researcher generated 50,000 fake statistics. Google buried it. ChatGPT cited it 10,000 times and laundered the lies into real reports.
Research · Source: Hacker News
What happened
Metehan Yesilyurt built a website called StateGlobe with 50,000 fake statistics. He generated the data in one day using GPT-4.1. The site claimed to offer data on SEO, e-commerce, and web technology across 200 countries. He spent fifty dollars total on the domain and generation. Google indexed it initially. Then two algorithm updates hit, and Google completely buried the site. It now gets almost zero impressions on traditional search.
ChatGPT did the exact opposite. Over seven months, the site racked up more than 10,000 AI sessions. In just the last thirty days, the ChatGPT-User bot fetched pages from the site 72,000 times. These were not training crawls. These were live queries where ChatGPT pulled fake data to answer real users.
The fake data leaked into the real world. An academic paper cited the made-up numbers. An economics publication in India and a major Qatari bank used the fake data in their official research. The site also gained over 330 referring domains from major brands like Shopify, Bluehost, and DHL. Yesilyurt later ran a second test at Peec AI using real data, and it immediately attracted links from AI-generated content sites.
Key facts
- 50,000 — Fake statistics generated for the experiment
- $50 — Total cost to build the site and generate the data
- 72,000 — Requests from ChatGPT-User bots in a single 30-day period
- 10,000+ — Sessions driven entirely by ChatGPT over seven months
- 330+ — Referring domains acquired including Shopify and DHL
Why it matters
Generative Engine Optimization is completely decoupled from traditional SEO. You can have zero Google visibility and still dominate AI answers. When an LLM writes an article, it breaks the prompt into sub-queries called fanout queries. It actively hunts for specific numbers to fill knowledge gaps. A page optimized for what models search will beat a page optimized for human keywords.
This exposes a massive security flaw called citation laundering. LLMs operate on a tight budget for latency and compute. They use rerankers to score relevance and freshness, but they do not verify truth. A hallucinated number gets cited by an LLM. A lazy writer publishes it. A bank cites the writer. That bank report then becomes high-authority training data for the next generation of models. The lie becomes permanent.
For builders
Build simple statistics pages for your niche
LLMs actively hunt for data to ground their answers. Publish single-claim statistics with clear sources and current years in the headings. Use server-rendered HTML instead of hiding data behind JavaScript. You will capture AI referral traffic while your competitors waste money fighting for dead Google rankings.
Do not trust LLM citations blindly
Engineers building research tools must add verification layers. If your application serves unverified LLM statistics to enterprise clients, you risk passing on fabricated data. Your users will lose money when they base business decisions on hallucinations. Fact-checking is expensive, but losing your enterprise contracts costs more.
Track retrieval bots in your server logs
Google Search Console is useless for measuring AI traffic. You must monitor your server logs or Cloudflare for agents like ChatGPT-User and PerplexityBot. Founders who rely on traditional analytics will fly blind. You need to track exact number quotes to find unlinked mentions of your proprietary data.
My take
We are poisoning our own well. I build AI products every day, and seeing how easily a fifty dollar script bypassed OpenAI grounding mechanisms terrifies me. If we do not fix retrieval validation, the next generation of models will train on a synthetic garbage dump.
Original reporting: Hacker News. This is my rewrite and opinion.