ABSTRACT: In unconscious semantic priming, an unidentifiable visually masked word (the prime) facilitates semantic classification of a following visible related word (the target). Three experiments provide evidence that masked primes are analyzed mainly at the level of word parts, not whole-word meaning. In Exp 1, masked nonword primes composed of subword fragments of earlier-viewed targets functioned as effective evaluative primes. (For example, after repeated classification of the targets angel and warm, the nonword 'anrm' acted as an evaluatively positive masked prime.) Exp 2 showed that this part-word processing was potent enough to oppose analysis at the whole-word level. Thus, smile functioned as an evaluatively negative masked prime after repeated classification of smut and bile. Exp 3 found no priming when masked word primes contained no parts of earlier targets. These results suggest that robust unconscious priming (a) is driven by analysis of part-word information and (b) requires previous classification of visible targets that contain the fragments later serving as primes. Contrary to a widely held view, analysis of subliminal primes appears not to function at the level of analysis of complete words.