The boom of online social media and microblogging platforms has rapidly alter the way we consume news and exchange opinions. Even though considerable efforts try to recommend various contents to users, loss of information diversity and the polarization of interest groups are still an enormous challenge for industry and academia. Here, we take advantage of benign social bots to design a controlled experiment on Weibo (the largest microblogging platform in China). These software bots can exhibit human-like behavior (e.g., preferring particular content) and simulate the formation of personal social networks and news consumption under two well-accepted sociological hypotheses (i.e., homophily and triadic closure). We deployed 68 bots to Weibo, and each bot ran for at least 2 months and followed 100 to 120 accounts. In total, we observed 5,318 users and recorded about 630,000 messages exposed to these bots. Our results show, even with the same selection behaviors, bots preferring entertainment content are more likely to form polarized communities with their peers, in which about 80\% of the information they consume is of the same type, which is a significant difference for bots preferring sci-tech content. The result suggests that users preference played a more crucial role in limiting themselves access to diverse content by compared with the two well-known drivers (self-selection and pre-selection). Furthermore, our results reveal an ingenious connection between specific content and its propagating sub-structures in the same social network. In the Weibo network, entertainment news favors a unidirectional star-like sub-structure, while sci-tech news spreads on a bidirectional clustering sub-structure. This connection can amplify the diversity effect of user preference. The discovery may have important implications for diffusion dynamics study and recommendation system design.