Zombie Stats: Where the Numbers AI Quotes Come From

The Reliability Gap in AI Statistics

AI assistants frequently provide confident numerical answers that lack empirical grounding, often citing secondary sources that repeat unverified claims. A study by Trakkr Research involving 200 common questions found that only 40% of cited statistics lead back to an original study, while 11% lead to no source at all, and 40% stop at dead-end or inaccessible pages.

The Anatomy of a Zombie Stat

"Zombie stats" are figures that persist in the digital ecosystem despite having no verifiable research behind them. In one documented case, the claim that "44% of salespeople give up after one follow-up" is widely cited by AI models. Tracing this number reveals a trail of blog posts and freelance articles that reference one another, but never point to an actual study. This cycle of citogenesis—where AI models aggregate and repeat unverified information from the web—creates a false sense of consensus, even when the underlying data is nonexistent.

Performance Across AI Models

While all major AI assistants struggle with source verification, their citation habits vary. ChatGPT demonstrated the highest sourcing rate at 57%, while Claude, Gemini, and Google AI Overviews ranged between 38% and 46%. Notably, a higher frequency of citations does not correlate with higher accuracy. The models often rely on company blogs, news roundups, or statistics aggregators rather than primary research, which explains why even when a citation is provided, it frequently fails to lead to the original data.

Why Numbers Shift and Fail

Even when a trail successfully leads to a study, the data is often distorted. Approximately 17% of the traced statistics had been altered in context, year, or scope during their journey through the internet. For example, a 2004 study regarding 15 women evaluating health websites was eventually transformed into a generalized claim about the design preferences of all website users. These shifts occur as information is stripped of its original context and repurposed in marketing materials or secondary reports.

Best Practices for AI-Generated Data

Users should treat AI-provided statistics as leads rather than definitive facts. To verify information, users should:

• Follow the trail: Click through citations until reaching the primary source. If a link leads to a listicle or a company blog, continue searching for the original research paper or government report.
• Verify context: Check the sample size, the demographic, and the year of the study. A number is only as valid as the methodology behind it.
• Ignore consensus: Do not assume that because multiple AI models provide the same figure, the statistic is accurate. Models often share the same training data and common web biases.

Community Perspectives

Discussion surrounding these findings highlights a broader concern regarding the reliance on AI for due diligence. Critics note that the "seductive" nature of AI-generated charts and confident, cited answers can discourage users from performing their own verification. Some observers argue that the problem is not unique to AI, noting that human-generated content has long suffered from the propagation of "made-up" statistics, but AI accelerates this process by providing an veneer of authority to unverified claims. As one commenter noted, "A citation tells you where the AI found a number, not where the number came from."