The depth at which parafoveal words are processed during reading is an ongoing topic of debate. Recent studies using the RSVP-with-flanker paradigms have shown that a semantically implausible word in a sentence elicits a more negative N400 component than a plausible one already before the word enters foveal vision. While this finding suggests that word meaning can is accessed in parafoveal vision and used to rapidly update the sentence representation, evidence of similar effects in natural reading situations is still scarce. In the present study, we combined the co-registration of eye movements and EEG with the deconvolution modeling of fixation-related potentials (FRPs) to test whether semantic plausibility is processed parafoveally during natural Chinese sentence reading. For one target word per sentence, both its parafoveal and foveal plausibility were orthogonally manipulated using the boundary paradigm. Consistent with previous eye movement studies, we observed a delayed effect of parafoveal plausibility on fixation durations that only emerged on the foveal word. Crucially, in FRPs aligned to the pre-target fixation, a clear N400 effect emerged already based on parafoveal plausibility, with more negative voltages for implausible previews. Once participants fixated the target, we again observed an N400 effect of foveal plausibility. Interestingly, this foveal N400 effect was absent whenever the preview had been implausible, indicating that when a word's (im)plausibility is processed in parafoveal vision, this information is not revised anymore upon direct fixation. Our results provide convergent neural and behavioral evidence for the parafoveal processing of semantic information in natural reading.