Yes, it can. When people talk about expressive power it’s related to theoretical guarantees, whether a given GNN model can distinguish non-isomorphic graphs.

However, the problem of distinguishing non-isomorphic graphs is different from the problem of node/graph classification, so there is not much you can say about the performance of the model for a particular problem. It’s rather to show that theoretically, it can be good. But it does not guarantee you anything for a particular dataset.