Publish fake stats, get cited in LLMs

What it claims

StateGlobe (stateglobe.com) is a fake statistics site that a Peec AI researcher built in one day in March 2026. Google stopped showing it shortly after launch, but ChatGPT kept retrieving and citing it. The author's conclusion is that ranking on Google and being retrievable by an LLM are separate problems. Retrieval systems reward specific, fresh-looking numbers and do not check whether those numbers are true. A second test with verified statistics on peec.ai/ai-search-geo-statistics also drew links quickly, which the author reads as evidence that the tactic does not depend on the data being fake.

How it works

The content came from GPT-4.1, using a prompt that was a list of countries and topics plus two sentences. It covered SEO, content marketing, social media, e-commerce and web technology across 200 countries. No number was checked. The 50,000 stats pages cost about $30, and 15,000 listicles added later cost about $20. Indexer tools got almost the whole site indexed within a week.

The page describes ChatGPT search this way:

flowchart LR
  P[User prompt] --> F[Fanout sub-queries]
  F --> S[Search sources]
  F --> C[OpenAI cache/index]
  S --> R[Reranker: relevance, not truth]
  C --> R
  R --> G[Open a few pages, ground answer]

Sub-queries add modifiers such as "statistics", "data", "report" or the year. Rerankers score relevance and do not check truth, and no fact-checking step exists because of latency and cost. Stats pages also put the year in the title, headings and URL, which the author says feeds ChatGPT's recency bias.

Evidence

• Google: two Google updates right after launch collapsed visibility. Indexed pages now get 5 to 20 impressions, and a subdomain test did not help.
• Bot traffic: Cloudflare showed 600,000+ bot requests in the last 30 days. About 72,000 came from ChatGPT-User, the real-time fetch agent, not a training crawler.
• Sessions and visibility: GA4 shows 10,000+ ChatGPT sessions since March. In Peec's own tracker, the site ranked second for visibility on a few tracked prompts on October 7, behind Statista.
• Links and citations: Ahrefs and Semrush count 330+ referring domains. Linked citations appear in an academic paper, investment research from Qatar's largest financial group, an Indian economics publication and a Canadian influencer community report. Unlinked use was not counted.

The author names the risk "citation laundering". Each institutional citation makes a fake number look sourced, and the number may eventually reach training data.

How to use it

The author's checklist for real stats pages:

• Write one self-contained claim per sentence, because retrieval works on passages.
• Put the source and year next to each number.
• Use headings that match fanout queries, such as "[Topic] statistics", "[Topic] market size" and "[Topic] adoption by country".
• Show an update date and keep it current.
• Serve the data as server-rendered HTML, not through JavaScript, tabs or images. Check robots.txt and CDN or WAF rules for blocked AI bots.
• Add original data where possible.

For tracking, Search Console does not show AI bots. The author tracks ChatGPT-User, OAI-SearchBot and PerplexityBot in server or Cloudflare logs, plus GA4 referrals from chatgpt.com and perplexity.ai. They also search their exact numbers in quotes to find unlinked use.

Limits

This is a single uncontrolled experiment on one domain, and the metrics come from Cloudflare, GA4, Ahrefs, Semrush and Peec's own tool, not independent measurement. The real-data test reports no numbers. The author notes that direct sessions were excluded as likely bot traffic, that some pages may fail to load, and that the gap may close now that it is published. The site is still live and now carries a disclaimer.