The Quality of Resources Used in Google AI Overviews for Health Topics

Background: The growing number of consumers seeking health information online are increasingly offered Google AI Overviews atop their searches. Recent research has identified pitfalls of using large language models for health information generation. Objectives: To examine sources employed in generating AI overviews for minority-language queries and evaluate their quality. Methods: Query topics were derived from the Czech National Medical Library's patient information portal, entered in Google, and information about sources linked to AI overviews was collected and statistically analysed. Results: Of 145 distinct queries, 40 (27.6%) did not yield AI overviews. The remaining 105 queries produced 1,231 unique URLs (mean=11.7). General web pages were most frequent (68.9%, n=848), followed by Wikipedia (5.4%, n=66) and peer-reviewed articles (2.9%, n=35). Private-sector entities owned 55.7% of sources, public sector 29.5%, and non-profits 14.8%. Quality scores ranged from −1.5 to 7.0 (mean=1.85). Poor-quality sources comprised 28.4%, while 10.6% were excellent. Discussion: Quality variability strongly correlates with sector ownership. Public sources and structured initiatives demonstrate best practices, while private entities often introduce commercial bias and non-evidence-based content. Conclusion: Source quality for Czech health AI overviews is highly variable and sector-dependent. Rigorous quality control measures are essential for consumer safety.