Geometry plays a vital role in 3D point cloud semantic segmentation since each category of object exhibits a specific geometric pattern. However, popular point cloud semantic segmentation methods ignore this property during feature aggregation. In this paper, we propose a novel Covariance-based Geometry Encoder (CGE) to learn latent geometry representation in point clouds and break this limitation. Specifically, we find that the classic covariance matrix can represent geometry implicitly in a point neighborhood, and we can learn geometry representation through simple multi-layer perceptrons to enhance the point features in a deep network. The proposed CGE module is generally applicable to any point-based network, while only requiring a little extra computing. Through extensive experiments, our method shows competitive performance on both indoor and outdoor benchmark datasets. Code will be publicly available.