Access to word meaning outside of awareness is generally accepted now, at least as indexed by masked priming. Yet, it is not clear what kind of information is grasped subliminally, and whether the mechanisms underlying conscious and unconscious semantic processing are the same. Here, we address  this issue by comparing the state-of-art Distributional Semantic Model (DSM), found to successfully predict a wide range of language-related phenomena, with Pointwise Mutual Information (PMI), a measure of association between words based on their mere co-occurrence in language use. In particular, we test whether the two metrics predict facilitation in a dataset comprising visible and masked primes. Subliminally, we find no significant effect, although participants with better prime detection show larger priming. Supraliminally, PMI outperforms DSM in the fit to the behavioral data. According to our results, semantic priming may mostly be explained in terms of association between words as reflected in their co-occurrence patterns.