Graphemic Output Buffer (GOB) disorder is defined as difficulty with the serial output of a sequence of letters in the output stage of the spelling process. In their rehabilitation study with a GOB patient, Sage and Ellis (2006) found that improvement on treated words generalised to untreated words from the same orthographic neighbourhoods as treated items, but not to other unrelated words. GOB patients frequently show a bow-shaped accuracy curve across letter positions, where letters at the middle positions are most error-prone. It may be that consistent letters at these middle letter positions across neighbourhoods modulate this effect. Spelling was treated using an Anagram and Copy Treatment (ACT) and generalisation to three untreated sets was examined: (1) neighbours of treated words with shared middle letters (e.g., clock–block), (2) neighbours with different medial position letters (e.g., clock–click), and (3) unrelated words (e.g., clock–puppy). Improvement was found for untreated neighbours with shared middle letters. There was no effect of training on the unrelated word set, and a negative impact on untreated neighbours with changed middle letters after the treatment. We attribute these results to top-down support from learned lexical representations, which facilitate spelling of untreated neighbours with shared middle letters but impede neighbours with changed middle letters. This latter result is attributed to interference from neighbours in the trained set strengthening competing letter representations at middle positions.