Phylogenies that include long, unbranched lineages can be difficult to reconstruct. This is because long-branch taxa (such as rapidly evolving species) may share character states by chance more often than more closely related taxa share derived character states through common ancestry1. Despite Kim's warning that added taxa can decrease accuracy2, some authors have argued that the negative impact of this error, called ‘long-branch attraction’, is minimized when slowly evolving lineages are included to subdivide long branches3,4,5. From this they have concluded that increasing the number of species sampled per lineage results in better accuracy than increasing the number of characters per species6. We find, using computer simulations, that adding characters can be the more favourable strategy, even for long-branched trees, and that adding slowly evolving taxa to subdivide long branches can reduce accuracy.