Web recommendation systems have become a popular means to im-prove the usability of web sites. This paper describes the architecture of a rule-based recommendation system and presents its evaluation on two real-life ap-plications. The architecture combines recommendations from different algo-rithms in a recommendation database and applies feedback-based machine learning to optimize the selection of the presented recommendations. The rec-ommendations database also stores ontology graphs, which are used to semanti-cally enrich the recommendations. We describe the general architecture of the system and the test setting, illustrate the application of several optimization ap-proaches and present comparative results.