Previous studies from different languages have suggested that online pronoun resolution is sensitive to semantic, syntactic and phonological information about the antecedent, but are challenged by the finding that, in English, reading times on pronouns vary inversely with the frequency of their antecedents. In two eye-tracking experiments, I re-examine this question by exploring how the lexical frequency of the antecedent affects reading times after a pronoun. The results show that reading times following pronouns are reliably predicted by the antecedent's lexical frequency: pronouns with infrequent antecedents were read more slowly than pronouns with frequent ones. Furthermore, fitting of the ex-Gaussian distribution (Ratcliff, 1979) to individual participants' RT data revealed identical profiles at the antecedent and post-pronoun regions. These results support the claim that processing a pronoun involves re-accessing the lexical entry of its antecedent. I contrast this lexical re-access account with a purely conceptual account, and I propose that the same cognitive processes triggered by a low frequency word at the antecedent region take place after a co-referential pronoun has been encountered.