The claim that visually masked, unidentifiable words are analyzed at the level of whole word meaning has been challenged by recent findings indicating that instead, analysis occurs mainly at the subword level. The present experiments examined possible limits on subword analysis. Experiment 1 obtained semantic priming from pleasant- and unpleasant-meaning subliminal words in which no individual letter contained diagnostic information about a word’s evaluative valence; thus analysis must operate on information more complex than that contained in individual letters. Experiments 2 (...) and 3 showed that analysis must operate on information more complex than that represented by individual bigrams or trigrams . These findings suggest that while subliminal priming is driven by subword analysis, the effective units of analysis are distributed widely across at least short words. (shrink)