Structuralists and generativists define word classes distributionally (Palmer 1971, Baker 2003, Aarts 2007), while cognitive linguists take a semantic (Langacker 1987a) or semantic-pragmatic approach (Croft 1991, 2001). Psycholinguistic research, by contrast, has shown that phonological properties also play a role (Kelly 1992, Monaghan <i>et al.</i> 2005). This study reports on a production experiment involving English nonce nouns and verbs. The data confirm the importance of phonology, whilst also suggesting that distributional facts are involved in lexical categorisation. Together with the existing psycholinguistic evidence, the results show that both the generative and cognitive models of word classes are too restricted. However, the usage-based model can accommodate the facts straightforwardly. This was already anticipated by Taylor (2002) but is worked out in more detail here by elaborating on his notion of phonological “sub-schemas” and by bringing together insights from Croft (1991, 2001) related to discourse propositional act constructions and recent suggestions by Langacker (2008b) concerning “summary scanning” and “sequential scanning”.