News · 2026-10-21

A researcher seeded 50,000 fake statistics and documented ChatGPT citations

Peec researcher Metehan Yesilyurt says he published 50,000 invented statistics and later documented ChatGPT retrieving and citing the resulting site. His October 7 account also shows third-party reuse, illustrating how a fabricated number can acquire an apparent citation trail. The evidence concerns selected observations and self-reported traffic, rather than a controlled comparison across chatbots.

Key facts

• Yesilyurt reports generating 50,000 invented statistics and later adding 15,000 listicles.
• The experiment began in March; Peec published the account on October 7, 2026.
• The author reports more than 72,000 ChatGPT-User fetches in the final 30-day measurement window.
• Primary source: Peec’s “Publish stats, get cited” account.

The experiment targets a failure that polished citations can hide. An assistant can retrieve a real page, quote it accurately and still deliver a false claim if that page invented its underlying data. The link proves that a statement was published somewhere. It does not, by itself, prove that anyone measured what the statement describes.

Yesilyurt says he built StateGlobe in March, used GPT-4.1 to generate the statistics, and spent about $50 on the overall exercise. He later expanded the site with list-based articles. His account documents continued visibility for selected prompts months after launch and presents examples of the invented figures being cited beyond the original site.

Peec calls this “citation laundering.” The phrase captures a change in appearances rather than a change in truth. A fabricated number appears on one page. Another publisher cites it. Later readers encounter the number inside a document carrying a more familiar institution’s name. Following only the latest link can make the claim look better sourced even though the chain still ends at the invention.

The analogy is a photocopied certificate. Passing the copy through several respectable offices does not authenticate the original signature. In the same way, a bank’s research document or an academic paper can repeat a weakly sourced claim without turning it into an independently measured fact. The provenance chain matters more than the prestige of its latest carrier.

Ground Truth’s retrieval-augmented generation lesson explains why this reaches assistants. Retrieval supplies outside material during an answer. It can reduce unsupported generation while importing errors from the retrieved source. A system that retrieves faithfully still needs a way to judge source quality and whether the cited material actually supports the claim.

The reported traffic is striking, but its meaning is limited. Yesilyurt reports more than 10,000 ChatGPT-attributed sessions since launch and more than 72,000 fetches in a 30-day window. A fetch shows a service requested content. It does not establish that the resulting answer endorsed a statistic, that a person read it, or that every request produced a distinct answer. Those outcomes require answer-level evidence.

The author does provide selected citation examples. What he does not provide is a neutral prompt population, a complete query log, a controlled comparison among named chatbots, or a reliable time-to-first-repetition measurement. The supported claim is that he observed retrieval and citation of the planted site, including continued visibility seven months after launch. Broader claims about all assistants or immediate propagation exceed that record.

His second test should also remain separate. In that exercise, he published real, checked statistics on a Peec page and says AI-written sites began linking to it. That observation concerns how statistics pages attract reuse. It does not independently replicate the fake-statistics result, and mixing the two tests would obscure which evidence supports which conclusion.

The Hacker News discussion raises the strongest methodological objection: retrieval or citation is not the same as a model accepting a claim as true. Commenters also ask whether requiring reputable sources would change the result. Those objections limit the conclusions without erasing the demonstrated risk that a fabricated source can enter an answer’s evidence chain.

Commercial incentives deserve visibility too. Yesilyurt works on generative search optimization, and Peec sells tools for measuring visibility in AI answers. The post uses its own dashboard and selected prompts. That is a reason to inspect the method and seek replication, rather than to dismiss the documented examples or treat them as a representative prevalence study.

The author raises possible propagation into training datasets but does not demonstrate it. This story is about observed retrieval, citation and third-party reuse. The practical response is to trace important numbers to their original measurement, look for a method and denominator, and distinguish corroboration from copying. A web of links becomes evidence only when at least one strand reaches something that was actually observed.

Primary source, verified: read the paper →

Ground Truth. (2026, October 21). A researcher seeded 50,000 fake statistics and documented ChatGPT citations. Ground Truth. https://groundtruth.day/news/peec-fake-statistics-citation-laundering-experiment.html

@misc{groundtruth:peec-fake-statistics-citation-laundering-experiment,
  title  = {A researcher seeded 50,000 fake statistics and documented ChatGPT citations},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/peec-fake-statistics-citation-laundering-experiment.html}
}

Topics: retrieval · misinformation · search · evaluation · data-quality