In a typical picture-word interference (PWI) task, picture-naming is faster/more accurate when the preceding prime words are identical to the picture name, but slower/less accurate when they are semantically related. However, ERPs time-locked to picture onset typically only show priming (facilitation) effects for both related prime types (smaller negativities compared to unrelated prime words). We replicated these effects and then conducted a novel behavior-contingent analysis for the semantic condition in which single-trial ERP data were re-averaged for word-picture pairs that either showed a facilitation pattern or an inhibition pattern (faster or slower RTs for related than unrelated pairs). The “facilitation bins” produced typical priming in early (200-400ms) and late (400-600ms) N400 epochs, while the “inhibition bins” produced reversed priming (greater negativity for related than unrelated prime words) in the late N400 epoch. In addition, after correcting for speech artifact, we observed reversed priming for the semantic condition in a later epoch (700-1200ms) that overlapped with the naming response (for both bin types). Overall, these results help resolve the apparent conflict between neural and behavioral responses in the PWI paradigm. Further, the pattern of results suggests that competition between lexical items can occur during both lexico-semantic activation and later speech production processes, lending some support to both lexical selection and response centered accounts of PWI effects.