Most computational models of visual word recognition assume that the positionof each letter within a word is perfectly encoded. This implies that the activation of the word JUDGE is predicted to be the same when the visual input is JUGDE or JUCRE. However, a number of experiments have shown that transposed-letter nonwords are more 'similar' to their base word than replaced-letter nonwords (eg, Perea & Lupker, 2003). Current computational models are unable to explain why transposed-letter similarity effects are so robust. Here, we propose a new model for encoding letter positions that can successfully deal these effects: the overlap model. The basic assumption is that letters in the visual stimulus have distributions over positions so that the representation of one letter will extend into adjacent letter positions. To test the model, we conducted a series of perceptual identification experiments. The overlap model produced very good fits to the empirical data. The model captures the fact that the position in the string of the transposed or replaced letters (external vs internal) has a large impact in the discriminability. The model also captures the fact that, compared to letter replacements, transposition of letters are harder to identify even when the transposed letters are up to three positions apart.