Trakkr asked five AI assistants 200 questions and traced 1,285 statistical citations through up to five pages. It found that 40% matched an original study, 8% changed a number or its meaning, and 40% could not be accessed or verified. A page containing a number was not necessarily its original source. This was a snapshot of a particular question set, and unverifiable does not mean false. Assistants can repeat the same citation chain, so consequential figures require checking the underlying population, date and definition.