Phishing attacks continue to be one of the most prevalent and damaging cyber threats, exploiting human vulnerabilities and advancing in sophistication. Despite significant progress in phishing detection techniques, new challenges arise as attackers adopt more dynamic and complex strategies. This survey explores the evolution of phishing detection methods, focusing on the machine learning and deep learning approaches that have emerged in recent years. We present an in-depth categorization of the feature types employed in these detection techniques. Additionally, we examine the critical role of modern feature representation techniques, including word embeddings, FastText, and BERT, in enhancing detection accuracy. A comprehensive taxonomy of phishing detection techniques is also presented, capturing the diversity of current approaches. Furthermore, we analyze the growing use of hybrid models and multimodal approaches, which combine multiple data sources to improve detection efficacy. Special emphasis is placed on the increasing need for real-time adaptation and the development of lightweight models suitable for deployment in resource-constrained environments, such as mobile devices and IoT platforms. Moreover, we identify key research gaps, including the lack of sophisticated defenses against emerging phishing tactics, challenges in dataset quality and availability, and the need for explainable and interpretable models. Through this comprehensive survey, we aim to provide a roadmap for future research in phishing detection, emphasizing the importance of continuous learning, dataset diversity, and hybrid detection strategies to counter evolving threats effectively.