In this paper, we describe an embedding-based entity recommendation framework for Wikipedia that organizes Wikipedia into a collection of graphs layered on top of each other, learns complementary entity representations from their topology and content, and combines them with a lightweight learning-to-rank approach to recommend related entities on Wikipedia. Through offline and online evaluations, we show that the resulting embeddings and recommendations perform well in terms of quality and user engagement. Balancing simplicity and quality, this framework provides default entity recommendations for English and other languages in the Yahoo! Knowledge Graph, which Wikipedia is a core subset of.