Fake statistics site kept getting cited by ChatGPT months after Google deindexed it

A $50 site of invented statistics

According to a write-up by Peec AI, a company that sells AI search visibility tooling, one of its team members spent roughly $50 to test how easily fabricated data spreads through AI-powered search. Work began in March 2026 on a brand-new domain, stateglobe.com, which presented statistics on SEO, content marketing, social media, e-commerce and web technology from 200 countries. All 50,000 statistics pages were generated in a single day with GPT-4.1, and every number was invented — the author says none were ever checked. A later batch of 15,000 listicle pages cost about another $20, bringing total spending to around $50.

The idea grew out of ongoing research on LLMs and grounding, the author says, crediting an August 2025 study by Seer Interactive along with work by iPullRank, Dan Petrovic and Wil Reynolds. The premise was simple: as more teams use LLMs to produce time-sensitive content, those models go looking for statistics to fill gaps in their knowledge, and a statistics page is exactly what they want to find.

Google buried the site almost immediately

The site initially indexed well and appeared in Google results within about a week. Then two Google updates arrived shortly after launch, and visibility collapsed. Peec reports that indexed pages have bounced between 5 and 20 impressions for months, and a subdomain created to test whether visibility could be recovered made no difference.

ChatGPT behaved very differently. Cloudflare crawl data for the 30 days before the write-up showed more than 600,000 bot requests, of which roughly 72,000 came from ChatGPT-User — the agent ChatGPT uses to open pages in real time while answering questions, not a training crawler. GA4 recorded more than 10,000 sessions from ChatGPT since launch. Ahrefs and Semrush attribute 330-plus referring domains to the site, including Shopify, Bluehost and DHL. On 7 October the site ranked second for visibility on a set of tracked prompts, trailing only Statista.

The fabricated numbers went mainstream

The invented data did not stay on the site. Peec reports that StateGlobe numbers were cited in an academic paper, used in investment research by Qatar's largest financial group, quoted by an economics publication in India, and included in a research report from a Canadian influencer community. Those are only the citations that linked back; the author assumes many more used the data without attribution.

Ranking and retrieval are different problems

The first lesson Peec draws is that Google ranking and LLM retrievability are separate concerns. When ChatGPT searches, it does not copy Google's top results: it rewrites a prompt into multiple sub-queries, pulls candidates from search sources and its own cache, reranks them, and opens a handful of pages to ground the answer. A page Google has dropped can perform well at every one of those steps, which is exactly what StateGlobe did. For anyone optimizing for AI search, Google rankings are one signal rather than the whole picture.

Citation laundering

The second lesson is more uncomfortable, and the author labels it 'citation laundering'. On day one, a fake statistic on a brand-new domain had no authority. Then an academic paper cited it, then a bank, and now the next researcher — or the next model — encounters a figure with a named source that serious institutions have referenced. The data never became true, but each citation made it look more trustworthy. Peec also warns of a longer tail: pages that are repeatedly retrieved and cited end up in web crawls, and web crawls feed training data, so a number that began as a GPT-4.1 guess could become something a future model simply knows.

The author emphasizes that the post is shared as research rather than a playbook for fake content, and argues that verifying sources should become a standard part of writers' and researchers' workflows.

Why it matters

The experiment is self-reported by a vendor with a commercial stake in AI search measurement, so the figures deserve some scepticism. But if the numbers are even roughly right, the finding is structural: conventional search quality signals suppressed a $50 fake-statistics site within months, while AI retrieval kept returning to it for half a year and helped move its invented numbers into academic, financial and journalistic contexts. Anyone consuming AI-generated answers — or publishing work built on them — now needs to treat citations as claims to verify against primary sources, because the citation graph that both humans and models rely on can be seeded cheaply and poisoned at scale.