The problem of inferring confidence sets of gene trees is discussed without assuming that the substitution model or the branching pattern of any of the investigated trees is correct. In this case, widely used methods to compare genealogies can give highly contradicting results. Here, three methods to infer confidence sets that are robust against model misspecification are compared, including a new approach based on estimating the confidence in a specific tree using expected–likelihood weights. The power of the investigated methods is studied by analysing HIV–1 and mtDNA sequence data as well as simulated sequences. Finally, guidelines for choosing an appropriate method to compare multiple gene trees are provided.