Phishing attacks and malicious URLs are growing threats in cybersecurity, making reliable detection methods more critical than ever. This study explores how SHAP-based feature selection and meta-ensemble learning can improve phishing URL classification. We apply SHAP (SHapley Additive ExPlanations) to identify the most crucial features for classification on ISCX-URL-2016 and the UCI Phishing Dataset. We developed a meta-ensemble model that combines XGBoost, SVM, and Random Forest as base learners, with Logistic Regression making the final decision. We performed the training on different sets of SHAP-ranked features (10 to 80 for ISCX, 10 to 30 for UCI) and measured the performance through accuracy, precision, recall, and F1-score metrics. Our findings indicate that while increasing the number of selected features generally improves classification performance, the benefits are none beyond a certain threshold. The best results achieved were an accuracy of 98.73% on ISCX-URL-2016 and 97.64% on the UCI Phishing Dataset, confirming the effectiveness of SHAP-guided feature selection.