A search researcher who filled a website with 50,000 entirely invented statistics says ChatGPT is still citing the fake numbers seven months later — even after Google effectively erased the site from its results. The experiment, detailed in a blog post by Metehan Yesilyurt, a GEO researcher at AI visibility firm Peec AI, shot to the front page of Hacker News on Tuesday, drawing hundreds of points and a lively comment thread.
The post, first published in early October, describes how Yesilyurt bought a brand-new domain, stateglobe.com, and used GPT-4.1 to generate 50,000 statistics pages in a single day — purportedly covering SEO, content marketing, social media, e-commerce, and web technology across 200 countries. Every number was made up. "I never checked a single one," he wrote. For more context on this story, see our ongoing AI news.
A $50 Disinformation Factory
The total cost of the experiment was roughly $50: about $30 to build and push the statistics site live on a basic AI web design template, plus around $20 more for 15,000 AI-generated listicles added later. The site was never intended as a real property — it was a probe into how modern AI-powered search decides what counts as a reliable source.
The inspiration came from longer-running research into LLM grounding, and in particular from an eye-opening study Seer Interactive published in August 2025. Yesilyurt's hypothesis was blunt: as publishers and marketers increasingly produce time-sensitive content with LLMs, those models must scavenge the web for seemingly authoritative data to fill gaps in their training knowledge — and statistics pages are exactly the format they reach for.
Google Buried It. ChatGPT Didn't.
The divergent treatment by the two platforms is the experiment's core finding. On Google, the site's fortunes collapsed almost immediately. After two search ranking updates following launch, visibility fell off a cliff, and indexed pages have bounced between roughly 5 and 20 impressions for months — the search giant effectively deindexed a site it appears to have flagged as low-value spam.
ChatGPT behaved differently. According to Cloudflare crawl data cited in the post, stateglobe.com received more than 600,000 bot requests in the last 30 days alone, of which around 72,000 came from ChatGPT-User — the agent ChatGPT uses to fetch pages in real time while answering a user's question. Those are not training crawls; each one represents ChatGPT pulling a fake statistics page to answer a live query. Over the site's seven-month life, Yesilyurt counts more than 10,000 AI sessions referred to the domain.
The Numbers Escaped Into the Real World
More troubling is where the invented figures traveled. Per Ahrefs and Semrush data cited in the post, the site has attracted links from more than 330 referring domains, including Shopify, DHL, and Bluehost. Yesilyurt writes that his made-up statistics have also ended up in an academic paper and in a bank's investment research — the exact downstream consequences he set out to demonstrate: AI answers get copy-pasted into decks, articles, and reports, laundering invented numbers into seemingly authoritative citations.
The site remained online and reachable at the time of writing, which means ChatGPT citations may still be accruing.
Not a Rigorous Study — By Design
The experiment has obvious limits, and Yesilyurt does not pretend otherwise. It is a single-site case study with self-reported measurements, the domain was promoted publicly as an experiment along the way, and the exact ranking factors Google used to bury the site cannot be isolated from the outside. The Hacker News discussion, where the post drew 284 points and more than 160 comments on Tuesday, quickly became one of the day's most-read threads. Yesilyurt also credits prior work on how ChatGPT's search index behaves, pointing readers to the ChatGPT Index research by a developer known as Tomek as context for why a deindexed-by-Google domain can remain firmly inside the index that ChatGPT's live retrieval actually draws from.
Why It Matters for AI Search
The experiment is a stress test of generative engine optimization — the fast-growing practice of tuning content to be cited by AI answer engines, known as GEO. Much of the GEO industry focuses on getting a brand into the AI answer; Yesilyurt's point is that the more consequential moment comes after, when users take that output and republish it.
It also lands amid mounting scrutiny of how AI assistants handle factual grounding. Search engines have spent two decades building spam defenses tuned to ranking signals; conversational assistants that retrieve pages on demand inherit those weaknesses without necessarily inheriting the defenses. A site that Google classifies as junk can still look like a statistics database to a retrieval system that mostly sees a clean page full of confident numbers.
For publishers, the takeaway is uncomfortable: AI-driven distribution can reward exactly the content that traditional search penalties would suppress. For users, the experiment is a reminder that a statistic with a specific-sounding source attached may have been invented for about the price of a takeaway dinner. And for the AI labs, it is one more data point that answer engines need stronger provenance and consensus signals before citation becomes the currency of the AI search era.
_This article is based on Peec AI's blog post by Metehan Yesilyurt and discussion on Hacker News. The claims described are the researcher's own reported measurements from Cloudflare, Ahrefs, and Semrush data._
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
