Decoding persuasive content traps: a comprehensive analysis of persuasive cues, topical uniformity, and engagement on YouTube’s recommendation networks

Abstract Users may be led into content traps by algorithmic recommendation systems, which are topically homogeneous, structurally cohesive areas that consistently reinforce similar narratives. A persuasion-aware framework for identifying such traps in YouTube recommendation networks is presented in this paper. Using three sociopolitical case studies (China–Uyghur discourse, Cheng Ho propaganda, and the 2024 Trump assassination attempt), we construct multi-hop recommendation graphs comprising 9748, 8489, 5553 unique videos and 14,307, 13,384, 9689 recommendation edges, respectively. We apply Focal Structure Analysis to extract structurally salient subnetworks, BERTopic to estimate weighted topical uniformity, and a large language model to extract persuasion cues grounded in four classical persuasion theories. In our study, we then associate group-level engagement metrics (views, likes, comments) with topical uniformity and persuasion cue intensity. While topical uniformity is negatively correlated with commenting activity, focal structures with higher topical uniformity concentrate more persuasive cues and draw greater engagement across datasets. These results provide a reusable framework for auditing persuasive risks in recommendation systems and describe content traps as simultaneously structural, semantic, and rhetorical phenomena.