40% of AI-Cited Statistics Trace to Original Source Study

Trakkr Research has exposed a troubling gap in AI assistant reliability: only 40% of statistics cited by major AI tools trace back to studies that actually support the quoted claim. The investigation, officially released on October 29, 2026, manually verified 1,285 citation trails across ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews, revealing that most numbers are repeated from secondary sources without verification of the original research.

Lead researcher Mack Grenfell hand-checked 150 citation trails and discovered that 40% stopped at sources readers cannot verify, including paywalled reports, deleted pages or content that never mentioned the statistic. Another 11% of citations led nowhere at all, ending at pages with no further attribution.

ChatGPT Leads in Citation Quality

Among the five assistants tested, ChatGPT achieved the highest verification rate at 57%, meaning more than half of its cited statistics traced to matching original studies. The model also quoted fewer statistics overall, averaging 1.3 numbers per answer compared to Claude's 6.8. Perplexity, Gemini, Claude and Google AI Overviews clustered between 38% and 46%, with margins too narrow to establish a clear ranking among the four.

The research methodology involved asking each assistant 200 questions requiring numerical answers, then following every citation link up to five pages deep. Where pages named sources without links, researchers manually searched for the original material. This rigorous approach uncovered systematic patterns in how AI tools handle factual claims.

Company Blogs Dominate Citation Chains

Only 21% of first-citation pages were original research. More than half (54%) were company blogs re-reporting figures without linking to source studies. When combined with news stories and statistics roundups, secondary sources accounted for the vast majority of AI citations. These pages repeat numbers from elsewhere, creating long citation chains that often never reach verifiable research.

The source type of the first citation significantly impacted verification success. Citations beginning at research firms or government pages reached matching studies 61% of the time. Company blogs and statistics roundups performed far worse, with only 37% and 35% respectively leading to verifiable sources. Even research firms fell short 38% of the time, frequently because full reports sat behind registration forms.

Zombie Stats Haunt All Five Assistants

Trakkr identified four zombie statistics, numbers repeated by all five AI assistants with no traceable research foundation. One example claimed 44% of salespeople give up after one follow-up attempt. Every assistant cited this figure confidently, but the citation trail ended at a freelance writer's portfolio sample from 2017 that listed no underlying study. These zombie stats persist because they sound plausible and get repeated across marketing content without verification.

Even when citations did reach original studies, 17% had changed along the way. While the specific digits moved in only one quarter of these cases, context often shifted dramatically. A 2004 study of 15 women evaluating health websites morphed into a claim that 94% of all first impressions of every website are design-related, losing crucial limitations about sample size, domain and methodology.

Consensus Provides No Accuracy Guarantee

Agreement among AI assistants did not correlate with citation quality. When researchers asked the same question to all five tools, they received four or five different answers on 28 of 71 occasions where all assistants provided percentages. The typical gap between the highest and lowest figure was 16 percentage points. On the 10 questions where all five agreed on the same number, only 2 of the 5 traced figures proved verifiable.

Trakkr validated its question sample by also analyzing 300 statistics from Google AI Overviews answers to 52,930 real everyday searches tracked between September 25 and October 25, 2026. The verification rate remained consistent at 40%, confirming the findings applied to actual user queries, not just researcher-designed test cases.

What This Means

This research reveals a fundamental challenge in AI-assisted information retrieval: confident presentation does not guarantee factual accuracy. Citations that appear authoritative often trace back to circular references, secondary reporting or dead ends. Users cannot distinguish reliable statistics from zombie stats by tone or citation style alone. As AI tools become primary research interfaces, the gap between cited claims and verifiable evidence demands urgent attention from developers, publishers and users who rely on these systems for factual information. Organizations using AI assistants for research should implement independent verification processes for any statistics that inform critical decisions.