This paper presents a single case study investigating the mechanisms underlying generalization of treatment benefits to untrained words in spelling. Brunsdon, Coltheart, and Nickels (2005) observed that untreated words that improved tended to be those whose errors were closest to being correct prior to treatment. These words also tended to be high in written frequency. The present study employed the same treatment techniques as those used by Brunsdon et al. with K.M., a developmental surface dysgraphic. During a first treatment the characteristics of words whose spelling improved without specific training were identified. These characteristics were then used in a second treatment to test whether it was possible to predict generalization. The results showed that treatment generalization to untreated irregular words was best predicted by neighbourhood size and frequency. We suggest that the processes underlying treatment generalization are based on the interaction between the orthographic lexicon and the graphemic buffer. Clinical implications are discussed.