Table 1.
Breakdown of the NewsGuard news sources dataset by country and reliability.
Table 2.
Volume of tweets by country and reliability.
Table 3.
Breakdown of the filtered dataset by country and topic.
Fig 1.
Topic modeling results on questionable and reliable news sources’ content across countries.
The size of each topic is given by the proportion of unique news sources contributing to it. The flows represent the interest shift of news outlets in different topics over time, and the size of each topic across years is proportional to the number of news sources discussing that topic.
Fig 2.
Cumulative number of the total retweets received by questionable sources vs the cumulative number of retweets for reliable sources across different countries and topics.
The x and y coordinates of each dot represent the cumulative number of retweets received by reliable and questionable sources respectively. This cumulative number is calculated by arranging the retweets based on their creation time and aggregating them on a daily basis. A linear trend suggests that the temporal patterns of questionable engagements mirror a scaled version of reliable engagements. Each graph displays the ρ coefficient, illustrating the ratio of volumes between retweets received by questionable and reliable posts.
Fig 3.
Similarity network among news outlets, where each news source is represented as a node, and edges represent audiences’ similarity among news outlets.
The color and shape of the nodes indicate the classification of the news source, and the thickness of the edges represents the level of similarity of retweeters between two news sources. We discarded edges with weights lower than the overall median of the edges. Each network represents the news outlets’ similarity on one topic for one country. Networks are represented using the Fruchterman-Reingold force-directed graph drawing algorithm.
Fig 4.
Analysis of users’ consumption behavior through retweets.
Each histogram represents the user count versus the fraction of news from questionable sources, ranging from entirely reliable (0) to entirely questionable (1). A dominant presence near lower fractions suggests a prevalent reliance on reliable sources. In contrast, significant increases near the higher end highlight segments influenced by questionable content.
Fig 5.
Community detection analysis of news outlets’ similarity networks.
Clusters were found using the Louvain clustering algorithm and sorted based on the percentage of questionable news outlets. The percentage of questionable sources in each cluster is color coded. Network edges with weights lower than the median value were discarded here, result with the complete network is reported in SI.