Pine & Lieven (1993) suggest that a lexically-based positional analysis can account for the structure of a considerable proportion of children's early multiword corpora. The present study tests this claim on a second, larger sample of eleven children aged between 1;0 and 3;0 from a different social background, and extends the analysis to later in development. Results indicate that the positional analysis can account for a mean of 60% of all the children's multiword utterances and that the great majority of all other utterances are defined as frozen by the analysis. Alternative explanations of the data based on hypothesizing underlying syntactic or semantic relations are investigated through analyses of pronoun case marking and of verbs with prototypical agent–patient roles. Neither supports the view that the children's utterances are being produced on the basis of general underlying rules and categories. The implications of widespread distributional learning in early language development are discussed.