The way in which idioms are processed and the nature of their underlying representations are subject to an ongoing debate. Most processing models agree that idioms are specific combinations of ordinary words. However, models differ with respect to the exact role that these words are allowed to play. In the present study we tested the hypothesis that the relations between idiom words are specified on a lexical processing level. Specifically, we tested whether the constituent words of an idiom activate each other in the absence of an idiomatic (phrasal) context. In two lexical decision experiments, we found that this is indeed the case. However, the effect is modulated by the type of target word that precedes the idiomatic targets. Targets that are literally related to the first idiom target prevent activation of the second idiom target. The results support the superlemma model of idiom processing (Sprenger, Levelt & Kempen, 2006).