A recommender systems (RSs) is a software application designed to suggest items of interest to users by analyzing their preferences, behaviors, or past interactions. One of the most important and popular methods in recommender systems (RSs) is collaborative filtering (CF), which makes suggestions for items on the basis of user preferences and similarity with other users or products. The similarity metrics used to find these associations have a significant impact on how well CF performs. In this paper a systematic comparative evaluation of six widely used similarity measures (Cosine, Adjusted Cosine, Pearson Correlation, Jaccard, Euclidean, Manhattan) and experimental analysis of several similarity metrics in both user-based and item-based CF are mentioned. We evaluated performance across multiple metrics (MAE, RMSE, Precision, Recall, and F1-score) to provide concrete, data-driven guidance. We explain the theoretical background of each measure and discuss how dataset sparsity and rating distributions influence metric suitability, and assess their effectiveness using two datasets: Food.com and Allrecipes.com. Our experiments show that cosine and Jaccard similarity work best for user-based CF. Cosine similarity demonstrates strong performance on both Allrecipes.com and Food.com datasets for user-based CF, while Jaccard also shows competitive results on Allrecipes.com. For item-based CF, cosine similarity consistently outperforms other metrics. These observations highlight the effectiveness of specific similarity measures in enhancing the precision of Collaborative Filtering-based recommendation systems.