Participants performed a 2-choice categorization task on visible word targets that were preceded by novel (unpracticed) prime words. The prime words were presented for 33 ms and followed either immediately (Experiments 1-3) or after a variable delay (Experiments 1 and 4) by a pattern mask. Both subjective and objective measures of prime visibility were used in all experiments. On 80% of the trials the primes and targets belonged to different categories (incongruent trials), whereas in the remaining 20% (congruent trials) they could be either strong or weak semantically related category members. Positive congruency effects (reaction times faster on congruent than on incongruent trials) were consistently found, but only when the mask immediately followed the primes, and participants reported being unaware of the identity of the primes. Primes followed by a delayed mask (such that participants reported being aware of their identity) produced either nonreliable facilitation or reliable reversed priming (strategic), depending on whether the prime-target stimulus onset asynchrony was either short (200 ms; Experiments 1 and 4) or long (1,000 ms; Experiment 4). Facilitatory priming with immediate mask was found strong (a) even for participants who performed at chance in prime visibility tests; and (b) for high but not for weakly semantically related category coordinates, irrespective of category size (animals, body parts). These findings provide evidence that unconscious congruency priming by unpracticed words from large stimulus sets critically depends on associative strength and/or semantic similarity between category coexemplars. (Contains 3 tables and 6 footnotes.)