Every AI assistant you ask for a number gives you one. It sounds confident, it comes with a citation, and it's usually wrong — not because the AI invented the figure, but because the figure was never real in the first place. Trakkr Research asked ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews 200 questions that call for a number, then followed 1,285 of the cited statistics back through the web, page by page, up to five clicks deep. The results are damning. Only 40% of traced statistics reached a study that actually says the same thing. Another 8% reached a study but the number or its context had shifted along the way — a 2004 study of 15 women judging health websites morphing into a universal claim about first impressions. The remaining 52% either hit a dead end (11% led nowhere at all) or stopped at a source that was locked, deleted, or simply silent on the number (40%). The AI didn't make the numbers up. The web did, and the AI faithfully repeated them. The source type matters enormously. When the AI's first citation pointed to a research firm or government page, 61-62% of trails reached a matching study. When it pointed to a company blog — which accounted for 54% of all first citations — only 37% got there. Statistics roundups fared even worse at 35%. The citation ecosystem is dominated by secondary sources that repeat figures without linking to origins. ChatGPT performed best, sourcing 57% of its statistics, but achieved this partly by quoting far fewer numbers: 1.3 per answer versus 6.8 for Claude. ChatGPT was also the only assistant that flagged when evidence behind a number was unclear. The other four assistants clustered between 38% and 46%, too close to meaningfully rank. More numbers means more opportunities to launder zombie stats. The disagreement problem is equally revealing. On 28 of 71 questions where all five assistants provided a percentage, four or five different figures appeared. The median gap between highest and lowest was 16 percentage points. On 31 questions, the spread exceeded 20 points. When asked what percentage of marketers use automation tools, answers ranged from 49% to 96%. Each answer arrived with the same confident tone. The core mechanism is what Trakkr calls 'zombie stats' — numbers that circulate through the web's citation chain with no study behind them. A freelance writer's sample article becomes a sales firm's stats page becomes an AI Overview's authoritative answer. The AI's citation tells you where it found a number, not where the number came from. This distinction is invisible to the user. The implications extend beyond any single assistant. The AI systems are faithful mirrors of a web where secondary sources vastly outnumber primary research, where numbers accumulate authority through repetition rather than verification. The 82% of the time the AI's cited page contained the figure is not reassurance — it means the laundering chain works. Only 21% of pages cited for a number were the original research.
