This research explores the performance of various Recommender Systems (RS) models, progressively increasing their algorithmic complexity, from simple baselines to neural networks, in both rating and ranking tasks. Specifically, using the 2023 public Amazon dataset, we implement and compare six models: Global Average (GA), Matrix Factorization (MF) and Neural Rating (a self-developed custom model) for rating predictions; and Most Popular (POP) Bayesian Personalized Ranking (BPR) and Neural Matrix Factorization (NeuMF) for ranking. Performance metrics include MAE, MSE and RMSE for rating; and AUC, NDCG and Recall@50 for ranking. Results show that MF and BPR consistently outperform simpler baselines in accuracy, while neural models offer competitive results, especially when user’s behavior is more uniform. The Neural Rating model also performs comparably to established models, mainly when it is properly tunned. Findings highlight that model complexity tends to yield better performance, but its effectiveness depends heavily on the characteristics of the dataset and its adaptation capability. This work reinforces the need for selecting RS models based on data structure and business goals and opens the door for future development of hybrid or more adaptive algorithms.