Trakkr traced 1,285 statistics cited by ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews back through citation chains up to five pages deep, finding only 40% reached an original study matching the claimed figure, while 8% reached a study where the number or context had mutated. The remaining 51% dead-ended: 40% stopped at unverifiable sources (paywalled, missing, or silent on the figure) and 11% led nowhere. AI models predominantly cited intermediary pages—company blogs (over half), news stories, and statistics roundups—rather than primary research (only 21% of citations). Starting from a research firm or government page yielded a matching study 61% of the time versus 37% from company blogs and 35% from roundups; even research firms failed 38% of the time, often due to gated reports. Among statistics reaching a study, 17% had altered en route: the digit changed in 26% of mutations, while context shifts dominated (measured variable 20%, year 17%, source 16%, population 14%). In a test of 71 questions where all five assistants provided percentages, 28 produced four or five different figures with a median spread of 16 percentage points; consensus did not indicate reliability—only 2 of 5 unanimous answers traced to a study, and 11 of 29 multi-assistant agreements did. ChatGPT cited fewest statistics per answer (1.3 vs. Claude’s 6.8) and performed best on sourcing rates. The authors recommend treating AI statistics as leads: follow citation chains to primary studies, verify sample, date, and population, and publish methodology alongside figures on open pages to prevent zombie stats—numbers endlessly recycled without empirical origin.