Study finds 16% of AI search engine citations come from AI-generated sources

Researchers Mowafak Allaham and Nicholas Diakopoulos from the University of Amsterdam audited four major generative search engines—ChatGPT, Microsoft Copilot, Google Gemini, and Perplexity—using 712 real-world queries spanning politics, health, and the environment. Their study, posted on arXiv (paper 2605.23684), reveals that roughly 16% of all cited sources across these platforms are themselves AI-generated content, not original human-authored material. This means when you ask a generative search engine a question, the citations it provides may point to synthetic articles or posts created by other language models, creating a troubling feedback loop.

The audit also uncovered that generative search engines tend to rely on a narrow set of frequently cited domains, while the vast majority of sources appear only once or twice across queries. This concentration raises red flags about diversity of information and potential bias. The authors warn that users may unknowingly treat AI-generated citations as equivalent to authoritative sources, increasing the risk of misinformation. The study adds to growing concerns about the reliability of AI-powered search and calls for better governance and transparency in how these engines select and cite web content.