Traditional customer churn prediction uses profile and transaction data, leaving unexplored textual features such as customer reviews. This work compares machine learning models for churn prediction that use conventional data with those that use reviews posted by customers about their purchases. Our experiments with the most used models for churn prediction in the literature reveal that, using conventional data, the models present the best performance with RFM segmentation, reaching up to 93% F1-Score. This value drops to less than 75% without RFM segmentation. In contrast, using BERT embeddings of review texts, an F1-Score of 96% is achieved.