Reading involves an interaction of linguistic and perceptual processes, in which top-down, lexical knowledge aids letter perception. Does such facilitation interact with letter co-occurrence regularities? Skilled readers were briefly exposed to strings of five consonants; critically, letters in position 2 and 4 were embedded in either high (L in e.g., GLVTZ) or low (e.g., NLRTZ) transitional probability (TP) triplets. When presented with two strings differing by the critical letter (GLVTZ vs. GRVTZ), participants correctly identified the right option more often in high-TP than in low-TP triplets. The effect disappears entirely when the task requires identifying single letters (Reicher-Wheeler task) and only survives for position 2 in a same-different task with response time constraints. This shows that letter identification is affected by letter co-occurrence statistics, but only when the task emphasizes string rather than letter processing, and coherently with a rapid deployment of spatial attention towards the beginning of letter strings. The genuine orthographic nature of these effects is also corroborated by a lack of facilitation with strings of non–alphabetic characters that share low-level visual features with letters.