Identifying letters and their relative positions is the basis of reading in literate adults. The Local Combinations Detector model hypothesizes that this ability results from the general organization of the visual system, whereby object encoding proceeds through a hierarchy of neural detectors that, in the case of reading, would be tuned to letters, bigrams, or other letter combinations. Given the increase of receptive fields by a factor of 2 to 3 from one neural level to the next, detectors should integrate information only for letters separated by at most 2 other characters. We test this prediction by measuring the impact of letter spacing on reading, purifying this effect from confounding variables. We establish that performance deteriorates non-linearly whenever letters are separated by at least 2 blank spaces, with the concomitant emergence of a word length effect. We then show that this cannot be reduced to an effect of physical size nor of visual eccentricity. Finally, we demonstrate that the threshold of about 2 spaces is constant across variations in font size. Those results support the hypothesis that the fast recognition of combinations of nearby letters plays a central role in the coding of words, such that interfering with this representation prevents the parallel analysis of letter strings.