In reading, it is often assumed that words are recognized sufficiently quickly, accurately, and unambiguously that downstream processes may proceed with perfect information about word identity. For example, word predictability is believed to affect early reading time measures, yet a word's predictability cannot be calculated without knowledge of the word's identity. We argue that such information is not, in general, available to the language processing system, and that it proceeds with only probabilistic information about word identity. We predict therefore that what have been analyzed previously as predictability effects must instead be based on noisy estimates of word predictability that are influenced by the predictability of visually similar words (neighbors). We test this prediction by building a Bayesian model of visual word recognition, using it to compute the `average neighborhood surprisal' of words in a corpus, and testing the ability of this novel measure to explain human reading time data.