60% of AI Statistics Have Unknown Sources? What Are "Zombie Statistics"?
機械翻訳 / Machine-translated
A
AI Friends
14時間前・17分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever asked an AI "What percentage of people do X?" and copied the number straight into a document? A study published in October 2026 by AI visibility service Trakkr found that the original research could not be verified for approximately 60% of such figures.
What tech media outlet GIGAZINE covered on October 30, 2026 was a study by Trakkr titled "Zombie Stats."
The finding: only 40% of statistics presented by AI could be traced back to the original research.
The breakdown of the remaining 60% is as follows:
In other words, only about 4 out of every 10 figures can be said to have solid backing.
Zombie statistics are numbers that keep being repeated even though their sources have changed, become outdated, or never existed in the first place.
Why don't they disappear? Because each webpage copies the numbers from the previous one. AI reads those copied pages and delivers the figures to countless users.
A prime example comes from the question: "What percentage of salespeople give up after just one follow-up?"
All five AI systems tested — ChatGPT, Perplexity, Gemini, Claude, and Google's AI Overview — answered "44%." With that level of agreement, it certainly seems credible.
However, tracing back the source cited in the AI Overview led to a 2020 sales company statistics page. That link, in turn, pointed to a 2017 sample article written by a freelancer — with no research cited at all.
Another well-known figure is "94% of first impressions are determined by design."
The original source was a 2004 study that analyzed comments from 15 women about why they rejected health-related websites. Over time, the claim morphed into something else entirely, and by 2026, AI Overviews were stating it as a universal rule applying to "any website."
The real problem isn't the number itself — it's that the original context and conditions quietly fall away.
According to Trakkr's official page, the methodology was as follows:
The percentage of statistics successfully traced to original research varied by AI:
ChatGPT scored highest and also produced the fewest statistics per response at 1.3. Claude, by contrast, produced the most at 6.8 per response. It seems reasonable to suggest that AIs generating more figures are more likely to include ones with weaker backing.
There's an easy misconception here. According to Trakkr, AI rarely fabricates figures outright — 82% of the pages cited actually contained the number in question.
The problem is that those pages were not the original research. Only 21% of cited sources included a link pointing back to the original study.
There were also differences by source type:
AI frequently cites corporate blogs, which have a low rate of traceable sourcing — this appears to be a major driver of zombie statistics.
The weakness of AI citations has been noted in other studies as well.
Ahrefs examined approximately 16 million URLs cited by six AI systems including ChatGPT. Traffic arriving via AI was 2.87 times more likely to land on a 404 page (a page that doesn't exist) than traffic arriving via Google Search.
A separate study by the European Broadcasting Union (EBU) and BBC found that 45% of AI-generated news responses contained significant errors. According to Journalism.co.uk, 31% of those errors involved sourcing issues.
Putting all three together:
All three converge on the same point: AI answers should only be used after verifying the sources.
This study was likely centered on English-language queries, but it's far from irrelevant in Japan — ChatGPT, Gemini, and Claude are all widely used here.
Consider a small business planner writing a presentation for next week's meeting and including the claim "customer acquisition costs 5 to 25 times more than retention." The AI answered confidently, so they included it without a second thought.
Yet according to Trakkr's research, even this figure — cited by all five AI systems — could not be traced to original research. When someone in the meeting asks "What's your source?", there may be no good answer.
The same applies to student reports, blog posts, and advertising copy. Statistics are persuasive precisely because they carry an air of authority — which makes the credibility damage all the greater when they turn out to be wrong.
Trakkr recommends three practices for anyone using statistics:
According to the study, most distortion happens not in the numbers themselves but in the surrounding context. Simply asking "Who was this figure actually about?" can be enough to spot a problem.
For those who publish statistics, Trakkr also recommends including methodology details — sample size, date, and a direct link to the source — alongside any figure you share. This is one of the most effective ways to ensure AI cites your work accurately.
This study has its own limitations, which the official page acknowledges:
Also worth noting: "unverifiable" does not mean "incorrect." It's best to treat these findings as the results of a single study conducted by Trakkr alone.
They are numbers that keep being recycled even though their sources have changed, become outdated, or never existed. Through repeated copying, they drift further and further from any original research.
According to this study, not really. The cited figure appeared in 82% of the pages AI referenced. The issue is that those pages themselves often weren't the original research.
In this study, ChatGPT came out on top at 57%. Even so, nearly half of its statistics couldn't be verified. Source-checking remains necessary regardless of which AI you use.
No. The 44% sales statistic was matched by all five AIs, yet no original research could be found. They may simply all be reading the same outdated page.
Open the link the AI provides, then follow the citations from that page. If you end up at a research firm or government agency's page, that's a relatively good sign.
The next time you get a number from an AI, try following the link two levels deep before putting it in your document.
This article is a cross-post from AI Friends.
