Google AI Overviews Cite Credible Sources But Make Unsupported Claims

An independent audit by researchers at Washington University in St. Louis has exposed significant discrepancies between Google AI Overviews and their cited sources, raising concerns about information accuracy and the economic sustainability of digital publishers. Conducted between March 13 and April 21, 2026, the study involved 55,393 searches across trending topics and will be presented at the October 2026 ACM Internet Measurement Conference. The research indicates that AI Overviews appear for 13.7 percent of all queries, with activation rates heavily influenced by query structure and topic. Nearly 65 percent of natural-language questions triggered an overview, compared to 9.5 percent of keyword-based searches. Longer queries also showed higher activation, while political and legal topics were significantly less likely to generate AI summaries, suggesting undisclosed editorial filtering by Google. When triggered, the system drew from an average of eight sources per response, with domains generally scoring higher on credibility metrics than traditional search results. Notably, nearly 30 percent of cited websites did not appear in the standard first-page listings, indicating a distinct source-selection algorithm separate from Google’s core ranking system. Despite the high quality of cited domains, the audit found substantial gaps between source material and the generated summaries. The team evaluated over 98,000 verifiable claims and determined that only 41.9 percent were fully grounded in the referenced text. Approximately 11 percent of claims were either unsupported or directly contradicted by the cited pages, a rate the researchers note is likely an upper bound due to limitations in capturing dynamic or user-generated content. Lead investigator Professor Jacob Montgomery emphasized that trusted citations do not automatically guarantee accuracy, urging users to verify AI-generated information independently. The financial implications for content creators are equally concerning. With AI Overviews delivering direct answers, publishers risk losing the traffic that traditionally fuels advertising, subscriptions, and sales. The study observed that more than half of the cited pages displayed ads, yet only 2.2 percent of overview results included Google-sponsored advertisements, with a small subset showing ads positioned above the AI summary. This dynamic suggests Google may be insulating its own ad revenue while redirecting user engagement away from external sites. Montgomery warned that without forthcoming licensing agreements or revenue-sharing frameworks, publishers reliant on advertising could face unsustainable losses over time. As search engines increasingly assume editorial roles, the findings underscore the urgent need for transparency in AI source attribution and equitable economic models for digital media. The complete methodology and datasets are available on the arXiv preprint server.