Theories of embodied semantics (ES) suggest that a critical part of understanding what a word means consists of simulating the sensorimotor experience related to the word’s referent. Some proponents of ES have suggested that sensorimotor activations are mandatory and highly automatic during semantic processing. Evidence supporting this claim comes from masked priming studies showing that unconsciously perceived spatial words (e.g., up, down) can directly modulate action performance on the basis of their meaning. However, a closer look reveals that such priming effects can be explained also in terms of symbolic (disembodied) semantic priming or nonsemantic mechanisms. In this study we sought to understand whether sensorimotor processing takes place during language understanding outside awareness. We used spatial words as a test bed and across 6 experiments we teased apart the possibility that action priming could be explained by: (a) nonsemantic mechanisms, (b) symbolic semantic priming, or (c) embodied semantic priming. The critical finding is that when symbolic and nonsemantic mechanisms were prevented, allowing only for a genuinely embodied semantic priming, no effect was found. Conversely, facilitation emerged in the same experimental paradigm when embodied priming was prevented and symbolic priming was allowed. Despite extensive testing, we found no evidence that unconsciously perceived words can activate sensorimotor processes, although these words are processed up to the semantic level. We thus conclude that sensorimotor activations might need conscious access to emerge during language understanding.