This document describes a new technique called "Latent Cross" for incorporating contextual data into recurrent neural network (RNN) recommender systems more effectively. The authors first demonstrate that modeling context as direct features in feed-forward neural networks is inefficient at capturing common feature interactions. They then apply this insight to design an improved RNN recommender system that uses Latent Cross. Latent Cross embeds the context and performs an element-wise product of the context embedding with the RNN's hidden states, allowing the model to better understand how context affects recommendations. The authors evaluate their approach on a large-scale RNN recommender system at YouTube and show that Latent Cross improves recommendation performance over conventional techniques.