By asahifile 
An SEO researcher built a website with 50,000 statistics. Every number was invented. Google buried it within weeks, but the fake figures kept travelling: seven months on, the site has drawn more than 10,000 visits from AI assistants, earned links from more than 330 referring domains including Shopify, Bluehost and DHL, and turned up in an academic paper and a bank’s investment research.
The experiment, run by Peec AI, is a useful stress test of how AI answers are assembled, and its results are uncomfortable. The pipeline that produced the citations is not exotic. It is the ordinary way models research a question.
Why statistics travel further than prose
The mechanism starts with how assistants break down a request. Asked to write about a topic, a model decomposes it into sub-queries and adds modifiers such as “statistics”, “data”, “report” and the current year. Pages built to match those queries get retrieved. Retrieval then operates on passages rather than whole documents, so a single self-contained sentence with a number in it can be lifted without the page’s context, caveats or disclaimer. Finally, the rerankers that order results score relevance, not truth: a passage that matches the query and looks authoritative is used.
Peec’s site was engineered for exactly that. One claim per sentence. A source and a year next to each figure. Headings that mirror the queries models run. A visible update date. Server-rendered HTML that needs no JavaScript to read. The numbers were fabricated; the packaging was impeccable. That is the point: citation-worthiness and accuracy are separate properties, and current retrieval rewards the first.
The laundering loop
The wider pattern is already documented under the name “LLM citation laundering”: an unverified claim gets repeated across dozens of independent-looking publications, those pages get indexed, and AI systems then cite the repetition as corroboration. One recent case involved a widely quoted claim that pages with FAQ schema are roughly three times more likely to appear in Google’s AI Overviews. The trail led back to a sponsored post on a search-industry site, with the cited supporting link returning a 404. When Ahrefs tested 1,885 pages against 4,000 controls across AI Overviews, AI Mode and ChatGPT, schema produced no meaningful citation increase, and a statistically significant 4.6 per cent decline in AI Overviews.
Each individual step in that chain is legible: an agency wants a quotable stat, a writer needs a source, a model needs a passage. The failure is emergent, and it is self-reinforcing. Once a number is cited widely enough, asking an assistant to verify it returns the same thirty articles, all tracing to the same untested origin.
What this means for anyone publishing numbers
The practical lessons cut both ways. The defensive one is that citation volume is now a poor proxy for reliability, and that a claim surviving an AI-assisted fact-check may only mean the claim has been copied often enough. The constructive one is that the traits which make a statistic citable are the same traits that make it useful to a human reader: one claim per sentence, an attributed source and year, a visible update date, and original data wherever possible. Peec’s author is explicit that fabricating figures is indefensible, that the site now carries a disclaimer, and that the tactic may stop working once models adapt.
There is also an incentive problem the experiment exposes. If retrieval rewards well-packaged numbers and ignores provenance, publishers who invest in verification compete on unequal terms with those who do not.
What to watch
- Whether AI search vendors ship provenance signals, such as citing a primary dataset rather than the blog that quoted it.
- Whether any academic publisher or bank that repeated the fabricated figures issues a correction, which would show the loop can be interrupted.
- Whether retrieval systems start down-weighting statistics pages that carry no linked primary source.
- Independent replication: the experiment is one operator’s data, and third-party tests would show whether the effect is general or site-specific.
Sources: Peec AI, The AI Search Brief on LLM citation laundering.