A growing literature suggests that readers generate predictions about various aspects of incoming linguistic input. Do expectations from context map onto lower-level expectations, and if so, how? We propose that comprehenders use internally generated predictions to explain the source of the input. In two experiments, we tracked eye-movements as subjects read sentences that generated strong (The boy saved the xxx) or weak (Mary had the word “xxx” tattooed …) expectations for nouns. Across contexts, subjects encountered target words (xxx) that had visual-form features that were either typical or atypical of nouns. In strongly predictive contexts, first-fixation and gaze duration measures were longer when the form of the target word was atypical with respect to the predicted category. In the less-biased contexts, no effect of word-form typicality occurred. These experiments provide eye-movement evidence that linguistic context is used to generate perceptual expectations about form-based properties of upcoming words during reading.