How faithfully AI Overviews represent source claims

AI Overviews misrepresent their sources at a measurable rate: decomposing responses to 55,393 queries (Mar–Apr 2026) into 98,020 atomic claims, an academic audit found 11.0% unsupported by the cited pages, with omission (missing context or qualifiers) the dominant failure mode rather than fabrication; source quality and claim fidelity were largely independent. The study is a preprint under review.