Abstract: In the age of personalized information communication, the social and cultural threat of information cocoons has drawn public attention. Under the personalized surface of “everyone has their preference”, whether falling into a “cocoon” leads users towards content convergence or differentiation is a controversial academic proposition and an urgent practical issue that needs to be tested. Based on the BERT model, 2 143 users randomly selected from Sina Weibo were analyzed, “information cocoons’ convergence” in social network user content was empirically studied. Results show that: (1) the degree of content similarity between any two users could be expressed and predicted by their information cocoon degree; (2) different users showed higher content convergence when the degree of their information cocoon rose, and showed lower content homophily when there was a big difference between their information cocoon degree. This result reflects on the cognition of the popular view of information cocoons from the perspective of similarity, and is conducive to understanding more deeply the complex mechanism and practical problems of social media users’ convergence of UGC (user-generated content).