Bootstrapping is a common method for assessing confidence in phylogenetic analyses. Although bootstrapping was first applied in phylogenetics to assess the repeatability of a given result, bootstrap results are commonly interpreted as a measure of the probability that a phylogenetic estimate represents the true phylogeny. Here we use computer simulations and a laboratory-generated phylogeny to test bootstrapping results of parsimony analyses, both as measures of repeatability (i.e., the probability of repeating a result given a new sample of characters) and accuracy (i.e., the probability that a result represents the true phylogeny). Our results indicate that any given bootstrap proportion provides an unbiased but highly imprecise measure of repeatability, unless the actual probability of replicating the relevant result is nearly one. The imprecision of the estimate is great enough to render the estimate virtually useless as a measure of repeatability. Under conditions thought to be typical of most phylogenetic analyses, however, bootstrap proportions in majority-rule consensus trees provide biased but highly conservative estimates of the probability of correctly inferring the corresponding clades. Specifically, under conditions of equal rates of change, symmetric phylogenies, and internodal change of <20% of the characters, bootstrap proportions of >70% usually correspond to a probability of >95% that the corresponding clade is real. However, under conditions of very high rates of internodal change (approaching randomization of the characters among taxa) or highly unequal rates of change among taxa, bootstrap proportions >50% are overestimates of accuracy.