News Source Citing Patterns in AI Search Systems
We address this gap by analyzing data from the AI Search Arena, a head-to-head evaluation platform for AI search systems. The dataset comprises over 24,000 conversations and 65,000 responses from models across three major providers: OpenAI, Perplexity, and Google. Among the over 366,000 citations embedded in these responses, 9% reference news sources. We find that while models from different providers cite distinct news sources, they exhibit shared patterns in citation behavior. News citations concentrate heavily among a small number of outlets and display a pronounced liberal bias, though low-credibility sources are rarely cited. User preference analysis reveals that neither the political leaning nor the quality of cited news sources significantly influences user satisfaction. These findings reveal significant challenges in current AI search systems and have important implications for their design and governance.
However, this gatekeeping role raises critical questions about which sources these systems prioritize and how their citation patterns shape public discourse (Memon and West 2024). These concerns are not new since audits of traditional search engines and news aggregators have shown that cited news sources concentrate among a relatively small number of popular outlets and exhibit liberal bias (Trielli and Diakopoulos 2019; Fischer, Jaidka, and Lelkes 2020). Built upon traditional information retrieval systems, AI search systems may amplify these issues through their complex and opaque mechanisms that can embed systematic biases in source selection (Yang and Menczer 2025).
Recent audits of popular AI search engines confirm these concerns. Using simulated user queries, researchers have shown that tools like Microsoft Copilot tend to favor mainstream news outlets (Brantner, Karlsson, and Kuai 2025). These systems have also been found to surface more lowcredibility sources than Google News and demonstrate leftleaning bias (Li and Sinnamon 2024; Jaidka and Furniturewala 2025). Given the contentious nature of the contemporary media landscape, AI systems’ tendency to favor sources with particular characteristics could contribute to information bubbles and exacerbate information disparities.
While this emerging evidence provides valuable insights, critical gaps remain. First, previous audits of AI search systems rely primarily on simulated user queries, which may not accurately reflect real-world usage patterns. Second, these audits typically examine only one or a few AI search systems, limiting their representativeness and preventing the identification of cross-model patterns. Third, these studies lack user preference data, making it difficult to determine which factors drive user satisfaction with search results.
In their qualitative study, Narayanan Venkit et al. (2025) recruit 21 expert users to compare AI search systems to traditional search engines. The experts commonly identify problems in the AI search systems, such as misattribution and misrepresentation of cited sources, missing citations for certain claims, and a lack of transparency in source selection. In particular, these experts express general distrust toward certain sources cited by AI search systems, especially forums, blogs, and opinion pieces.