Ratings indicate how favourably readers evaluated a book but provide limited information about which book characteristics contributed to that evaluation. This study investigated whether interpretable, review-derived cues intended to represent appeal-related book characteristics improve personalised Goodreads rating prediction beyond ratings-based information, book metadata, and Goodreads shelf labels. Using the Goodreads Book Graph, the analysis evaluated held-out ratings from 50,000 users in a warm-start setting in which evaluated users and books were already represented in the training ratings. Reviews were summarised using 19 review-normalised lexical cue rates intended to represent appeal-related characteristics such as pacing, characterisation, tone, prose, complexity, dialogue, and narrative form. User-specific associations with these cues were estimated from training-rating residuals. Adding the review-derived correction to a model already containing ratings, metadata, and shelf information reduced mean absolute error by 0.000098, corresponding to approximately 0.015%, while root mean squared error increased slightly. The same review correction produced a larger, although still small, MAE reduction of approximately 0.142% when added directly to the ratings-based additive-bias model. Randomly reassigning review corrections among observations with available review features did not reproduce the observed MAE reduction. These findings suggest that the implemented review-derived cues captured some personalised predictive information but provided little incremental predictive value once ratings-based information, book metadata, and shelf labels were already available.