A word's predictability, as measured by its cloze probability, has a robust influence on the time a reader's eyes spend on the word, with more predictable words receiving shorter fixations. However, several previous studies using the boundary paradigm have found no apparent effect of predictability on early reading time measures when the reader does not have valid parafoveal preview of the target word. The present study directly assesses this pattern in two experiments, demonstrating evidence for a null effect of predictability on first fixation and gaze duration with invalid preview, supported by Bayes factor analyses. While the effect of context independent word frequency is shown to survive with invalid preview, consistent with previous studies, the effect of predictability is eliminated with both unrelated word previews and random letter string previews. These results suggest that a word's predictability influences early stages of orthographic processing, and does so only when perceptual evidence is equivocal, as is the case when the word is initially viewed in parafoveal vision. Word frequency may influence not only early orthographic processing, but also later processing stages.