Nowadays, recommendation systems are a central component in a wide range of online services. One major challenge for them is to generate suitable recommendations for individual users who differ, among other things, in their interests and preferences. In this paper, we propose a deep learning-based product-to-product recommendation system that uses product embeddings to incorporate content-based product information and user data in form of historical checkouts. On top of that, we add a filter that rearranges the models' recommendations based on the individual user preferences to a product category that is derived from past activities. In order to evaluate and show the impact of our approach, we conduct tests on historical data and an A/B-test in a real online service that offers thousands of different customer packaged goods (CPG) to millions of users. The results show that our approach delivers suitable product recommendations to users and outperforms the current system in use.