This study examines the effectiveness of the Gravitational Search Algorithm (GSA) in feature selection for customer churn prediction models. Using access log data of an electronics retail e-commerce (EC) site in Japan (February–August 2024), models constructed with GSA-selected features were compared to those using all features. Nine machine learning models, including CatBoost, SVM, and Naïve Bayes, were evaluated using metrics such as AUC, accuracy, precision, recall, and F1-score. Boosting models, particularly CatBoost, demonstrated the highest performance with negligible differences between the two feature sets. In contrast, Decision Trees and Naïve Bayes showed improved metrics with GSA-selected features. GSA often reduced feature sets by approximately half, balancing performance and computational efficiency. However, it did not always select the most important features identified by full-feature models, suggesting its utility in addressing multicollinearity and redundancy. Future work involves feature expansion, such as integrating POS data, and improving GSA through parameter optimization and alternative encoding techniques. This research highlights GSA’s potential as an adaptable and effective feature selection method, contributing to improved predictive performance in EC customer churn scenarios.