Fake Statistics Site Kept Winning ChatGPT Citations After Google Buried It
A $50 site of 50,000 invented statistics kept drawing ChatGPT citations and 330+ referring domains after Google deindexed it, showing retrieval can reward fabricated numbers; the researcher warns the tactic may not last.
Peec AI researcher Metehan Yesilyurt seeded stateglobe.com with 50,000 fabricated statistics generated by GPT-4.1 in one day for about $30, then added 15,000 listicles for $20. Google buried the site after updates, but ChatGPT kept retrieving it: 600,000 bot requests in 30 days, 72,000 from ChatGPT-User, 10,000+ sessions, and 330+ referring domains including Shopify, DHL, and Bluehost.
The invented numbers were cited in an academic paper, investment research by Qatar's largest financial group, an Indian economics publication, and a Canadian influencer report. Yesilyurt attributes the effect to LLM fanout queries seeking statistics, rerankers scoring relevance rather than truth, and freshness signals.
A follow-up Peec page using checked, cited statistics also attracted links from suspected AI-written sites.
The bits that matter
- The experiment cost about $50: $30 for 50,000 statistics pages generated by GPT-4.1 and $20 for 15,000 listicles. The site claimed to cover SEO, content marketing, social media, e-commerce, and web technology across 200 countries, but every number was invented.
- Google visibility collapsed after two updates, leaving indexed pages at 5 to 20 impressions. Cloudflare data showed more than 600,000 bot requests in 30 days, including 72,000 from ChatGPT-User, and GA4 recorded 10,000+ sessions from ChatGPT.
- The fabricated numbers were cited in an academic paper, investment research by Qatar's largest financial group, an Indian economics publication, and a Canadian influencer community report. Each citation made the numbers look more trustworthy, a pattern Yesilyurt calls citation laundering.
- Peec AI then published a statistics page using checked, cited data. Suspected AI-written sites linked to it almost immediately, suggesting the tactic can work with accurate numbers, though Yesilyurt says it should never be abused.
- Recommended practices include one claim per sentence, source and year beside each number, headings matching fanout queries, visible update dates, server-rendered HTML, and checking robots.txt or CDN rules. Track AI bot logs, GA4 referrals, citations, referring domains, and unlinked mentions.
Discussion analysis
Analysis details
17 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 17 of 23 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
Retrieval scores relevance, not truth
LLM search fetches pages, pulls passages, and reranks for relevance; a specific number looks like evidence and can be used even when invented. Source context is often lost, so a seller's marketing copy can read like a textbook. Proposed mitigations include prompting for fact-checking, classifiers, or Google-style quality filtering, but each faces cost or incentive limits. If false claims enter model weights or retrieval caches, web takedown channels may offer no clear recourse.
Cheating pays in AI retrieval
The semantic web and agent-readable labels fail because there is a reward for cheating; scammers and counterfeiters mislabel, and verifying billions of labels is insurmountable. Adversarial pollution, flooding the web with fake math or bad data to raise AI costs, is floated as a countermeasure, but critics call it net negative for humanity and internet poisoning.
Google's quality filtering as defense
Google's ranking and spam systems are the primary existing defense, built with billions in investment, and may make its products less vulnerable than AI search. But ChatGPT retrieval often surfaces pages from the top Google results, so a polluted index can propagate into model answers. The same labeling and trust problems exist on the human web, though context helps humans judge sources.
Misinformation burden falls on users
The concern is that ordinary users cannot reliably distinguish misinformation from reality, and that accelerationists may accept a polluted information environment. Lack of due diligence is exacerbated by AI, and the result is a continuation of a bad trend.
Self-training feedback loop disputed
One view says the 2022 picture of LLMs regurgitating the web is outdated and that OpenAI has strong incentive to train classifiers distinguishing generated text. A counter-observation is that ChatGPT still cites pages from the first 5-10 Google results, suggesting retrieval remains dependent on web ranking.