This paper applies a computerized method related to that of glottochronology and addresses the question whether such a method is useful as a heuristic for identifying deep genealogical relations among languages. We first measure lexical similarities for pairs of language families that are normally assumed to be unrelated, using a modification of the Levenshtein distance as our similarity measure. We then go on to study how the similarities are statistically distributed. The average similarity is slightly greater than zero, suggesting a small effect of sound symbolism. The upper tail of the distribution extends to similarities comparable to what is typically found for well-established families or highest-order subgroups of old families, but the pairs of unrelated families with the highest similarities contain only a few languages. We conclude that the method may work as a useful heuristic, provided that the number of languages compared is taken into account.