Much of the evidence for morphological decomposition accounts of complex word identification has relied on the masked-priming paradigm. However, morphologically complex words are typically encountered in sentence contexts and processing begins before a word is fixated, when it is in the parafovea. To evaluate whether the single word-identification data generalize to natural reading, Experiment 1 investigated the contribution of morphological structure to the very earliest stages of lexical processing indexed by preview effects during sentence reading in the gaze-contingent boundary paradigm. Preview conditions systematically assessed the impact of prefixed and suffixed nonword previews that manipulated stem and affix overlap, and affix status, against an orthographically legal control baseline. Initial fixations on suffixed target words showed a preview benefit from nonwords that combined the target stem with a legitimate affix, but not with a nonaffix, whereas prefixed targets only benefited from an identical preview. When presented in a masked prime lexical-decision task in Experiment 2, the same stimuli yielded equivalent stem priming from suffixed and prefixed primes regardless of affix status, consistent with previous masked priming studies using similar nonword primes. The early effects of morphological structure selectively observed on parafoveal processing of suffixed words are inconsistent with recent nonmorphological, position-invariant accounts of embedded stem activation. These results provide the first evidence of morphological parafoveal processing in English and contribute to recent evidence that readers extract a higher level of information from the parafovea during natural reading than was previously assumed.