AI assistants cite unreliable statistics far more often than users may realize

Statistics generated or repeated by AI assistants can travel from obscure webpages into academic papers, business research and widely used online content, two new reports suggest. One experiment shows how easily a site filled with invented figures can attract AI traffic and citations. A separate study finds that fewer than half of the statistics cited by major AI tools can be traced to an original study that supports the claim.

The findings highlight a growing problem for anyone using generative AI for research, writing or marketing. A citation in an AI answer may show where a model found a number. It does not necessarily show where the number originated, whether it is current, or whether it is accurate.

In an experiment described by Peec AI’s Metehan Yesilyurt, a newly created site called StateGlobe published 50,000 AI-generated statistics on subjects including marketing, e-commerce and social media. Yesilyurt says he did not verify any of the figures. The site later added 15,000 list-style articles.

Google largely removed the site from search visibility, according to Yesilyurt. But ChatGPT continued to retrieve its pages. He reports that Cloudflare logs recorded roughly 72,000 requests from ChatGPT-User in one 30-day period, while Google Analytics showed more than 10,000 ChatGPT referral sessions over seven months. The site also gained links from more than 330 referring domains, he says.

Yesilyurt says some invented figures appeared in an academic paper, investment research and other publications. The example demonstrates what he calls a form of citation laundering. Once a fabricated claim is cited by a more established source, it can appear more credible to later readers, publishers and AI systems.

AI citations often lead to copies, not research

Trakkr Research examined the issue at a larger scale. Its researchers asked ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews 200 questions likely to produce numerical answers. They then traced 1,285 cited statistics through up to five links in the source chain.

The study found that 40 percent of the statistics led to a study that matched the claim. Another 8 percent reached an original source, but the number or its context had changed. Forty percent could not be checked because a source was inaccessible, missing the figure or otherwise unverifiable. Eleven percent led nowhere.

ChatGPT had the strongest result in Trakkr’s sample, with 57 percent of its cited statistics reaching a matching source. It also used the fewest figures per answer. The other four systems ranged from 38 percent to 46 percent.

The core problem was not usually pure invention by the AI, Trakkr says. In 82 percent of cases, the first page cited by an assistant contained the statistic. But only 21 percent of those first citations were original research. More than half were company blogs, while news articles and statistics roundups made up much of the rest.

Repeated figures can also lose critical context. Trakkr traced a common claim that 94 percent of website first impressions relate to design. The underlying 2004 research examined comments from 15 women assessing health websites, rather than all website users or all first impressions. Over time, later pages broadened the claim before AI systems repeated it as a general fact.

What content professionals should do

Users should treat an AI-provided statistic as a starting point, not a final source. Open the cited page, then follow its links until reaching the original study or dataset. Check the publication date, sample size, location and population studied. If the cited page is a blog post or statistics roundup, it is especially important to continue tracing the claim.

Publishers can reduce the same risk by placing source links, methodology and dates directly beside every statistic. Clear sourcing helps human readers and makes it easier for retrieval systems to identify the original evidence rather than a later copy.

Sources

• Publish fake stats, get cited in LLMs – Peec AI
• Zombie stats: Where the numbers AI quotes come from – Trakkr Research