Recommender systems are a constant research topic because of their large number of practical applications. There are many approaches to address these problems, one of the most widely used being collaborative filtering, in which in order to recommend an item to a user, data of other users` behaviors are employed. However, collaborative filtering algorithms do not always reach levels of precision required for the use in real applications. Within this context, the present work aims to evaluate the performance of the probabilistic latent semantic analysis (PLSA) applied to recommender systems. This model identifies groups of users with similar behaviors through latent attributes, allowing the use of these behaviors in the recommendation. To check the effectiveness of the method, there were presented experiments with problems of both web ad recommending and film recommending. An improvement of 18,7% were found in the accuracy of the recommendation of ads on the web and we also found 3.7% of improvement in Root Mean Square Error over the Means of Means baseline system for the Netflix corpus. Apart from the aforementioned experiments, the algorithm was implemented in a flexible and reusable way, allowing its adaptation to other problems with reduced effort. This implementation has also been incorporated as a module of LearnAds, a framework for the recommendation of ads on the web.