Methods of ancestor reconstruction are important tools for evolutionary inference that are difficult to test empirically because ancestral states are rarely known with certainty. We evaluated reconstruction methods for continuous phenotypic characters using taxa from an experimentally generated bacteriophage phylogeny. Except for one slowly evolving character, the estimated ancestral states of continuous phenotypic characters were highly inaccurate and biased, even when including a known ancestor at the root. This error was caused by a directional trend in character evolution and by rapid rates of character evolution. Computer simulations confirmed that such factors affect reconstruction of continuous characters in general. We also used phenotypic viral characters to evaluate two methods that attempt to estimate the correlation between characters during evolution. Whereas a nonphylogenetic regression was relatively inaccurate and biased, independent contrasts accurately estimated the correlation between characters with little bias.