The use of Graph Neural Networks (GNNs) in today’s recommender systems is a relatively new phenomenon, because of the way its structure allows capturing high-order connectivity and sophisticated relational patterns in user-item interaction graphs. Though many GNN architectures, such as GraphSAGE, GAT, GraphTransformer, GINConv, GCN, Graph Contrastive Learning, and AdaptiveGraphConv, have been very successful, their effectiveness in a variety of recommendations varies significantly. This work provides an evaluation of these seven GNN architectures on four heterogeneous recommender datasets: Anime Recommendations Database, RetailRocket E-commerce, Steam Video Games, and MovieLens 1M. Largescale experiments demonstrate different model abilities across domains. AdaptiveGraphConv and GraphSAGE are very robust in dense interaction settings, and attention-based models like GAT are competitive on graphs with features.