The comparative method depends crucially on the phylogenetic tree of the languages under comparison, but in many linguistic families, including Indo-European, the true tree is unknown. To circumvent this issue, frequency heuristics have been devised to enable comparative reconstruction over consensus trees. These heuristics come in different forms, but they are all based on the same methodological principle: if the number of homologous elements (e.g., lexical cognates) in the daughter languages meets a minimum threshold (canonically three), their ancestor is reconstructed to the root of the tree. In this paper, I demonstrate that frequency heuristics are not only unreliable but fundamentally misguided. As an alternative, I present a Bayesian method for inferring ancestral states that accounts for phylogenetic uncertainty by estimating the probability of a character state over a set of resolved phylogenetic trees.