Graph heterophily, also called disassortativity, is a property of networks that models the likelihood of nodes with different characteristics being connected. Several real-world graphs present high levels of heterophily, such as computer networks, where clients usually connect to servers, and dating applications, where users tend to connect with users of the opposite gender. Despite the relevant contexts in which heterophilous graphs are observed, current approaches fail to map nodes to meaningful low-dimensional embeddings. Indeed, Graph Neural Networks (GNNs), currently the state-of-the-art models for machine learning on graph-structured data, implicitly assume homophily, leading to worse performances on disassortative networks. Notwithstanding several works defining novel GNN architectures for heterophilous settings, there is still insufficient understanding of the relation between GNN performances and graph disassortativity.