Phishing attacks have become increasingly sophisticated, and their detection remains a persistent challenge in cybersecurity. This study presents a comparative evaluation of textual and numerical feature representations for phishing detection. We evaluated two independent datasets: the first contains 10,000 pre-processed records with 48 numerical attributes, while the second includes over 18,000 labeled email message bodies analyzed using natural language processing techniques with class weighting to address imbalance. A total of eight models—six machine learning algorithms (e.g. Random Forest, XGBoost, CatBoost) and two deep learning architectures (e.g. BERT and TabNet)—were trained and tested. Performance assessment based on accuracy, precision, recall, and F1-score, indicates that models using numerical features achieve superior detection, TabNet with recall and F1 values exceeding 99% compared to BERT with 98% for textual models. These results suggest that high-dimensional numerical features provide a more efficient basis for phishing classification than text-only representations.