Species tree estimation in the presence of incomplete lineage sorting (ILS) is a major challenge for phylogenomic analysis. Although many methods have been developed for this problem, little is understood about the relative performance of these methods when estimated gene trees are poorly estimated, owing to inadequate phylogenetic signal. We explored the performance of some methods for estimating species trees from multiple markers on simulated datasets in which gene trees differed from the species tree owing to ILS. We included *BEAST, concatenated analysis and several 'summary methods': BUCKy, MP-EST, minimize deep coalescence, matrix representation with parsimony and the greedy consensus. We found that *BEAST and concatenation gave excellent results, often with substantially improved accuracy over the other methods. We observed that *BEAST’s accuracy is largely due to its ability to co-estimate the gene trees and species tree. However, *BEAST is computationally intensive, making it challenging to run on datasets with 100 or more genes or with more than 20 taxa. We propose a new approach to species tree estimation in which the genes are partitioned into sets, and the species tree is estimated from the resultant ‘supergenes’. We show that this technique improves the scalability of *BEAST without affecting its accuracy and improves the accuracy of the summary methods. Thus, naive binning can improve phylogenomic analysis in the presence of ILS.