Trakkr research reveals low citation accuracy for statistics in AI

Analytics firm Trakkr Research conducted a verification of the data cited by popular AI assistants. The results showed that only 40% of the statistical figures provided in responses from ChatGPT, Perplexity, and Google AI Overviews are backed by real scientific or research evidence.

The testing revealed that models often generate "zombie statistics"—data that looks convincing but does not lead to a primary source. Among the systems analyzed, ChatGPT demonstrated the best performance, supporting 57% of its claims with links to actual studies. The other tools showed significantly lower accuracy in fact verification.

For users, this means it is essential to double-check any quantitative data obtained from neural network models. The high probability of hallucinations when dealing with numbers makes current versions of AI assistants unreliable sources for preparing analytical reports or making decisions that require reliance on accurate statistical indicators.