Phonological rules change the surface realization of words. Listeners undo these changes in order to retrieve the canonical word form. We investigate this so-called compensation for a French deletion rule, i.e. liquid deletion. This rule optionally deletes the final consonant of a word-final obstruent-liquid cluster. It can apply both before consonants and before vowels, but its application is about twice as frequent before consonants. Using a word detection task, we find an overall relatively low rate of compensation, which we argue is due to the relatively high perceptual salience of the rule. We also observe a clear effect of context, though: listeners compensate more than twice as often for a deleted liquid before a consonant than before a vowel. This is evidence that compensation involves fine-grained knowledge about the probability of the rule’s application in different contexts.