An AI-driven system has generated over 600 stories, adaptively levelled to reader proficiency for extensive reading, initially targeting first-year university students. Linguistic complexity is adjusted at the point of generation rather than selected from a fixed corpus, allowing us to compare predicted difficulty with actual student reading behaviour. The system collects fine-grained, page-level interaction data alongside learner comments and ratings, including time on each page, stop points, and total completion. Data from over 20,000 reading sessions are analysed using behavioural features such as completion rate, speed consistency, and re-reading frequency. Using these indicators, this study examines which linguistic or narrative features of stories sustain reading, as well as specific sections that delay, disrupt, or deter progress. Elevated reading speeds suggest superficial interaction, while reduced reading speeds may indicate increased cognitive load, but there are various intrinsic or extrinsic reasons why reading speed may change, from getting a coffee to not actually reading. Completion at a stable pace indicates their reading is comprehensible and compelling. Sentiment analysis of learner comments identifies patterns associated with successful and problematic texts. These findings are examined against intended text levels, with particular attention to performance at the lower and upper ends of the proficiency range.