Over the last decade, the study of multi-word units has become increasingly popular and now these units seem to have reached a status where they cannot be ignored. This paper should be seen as a recapituation of the discussion around multi-word units, especially focusing on corpus evidence of such units and what is perceived to be relevant findings. An unsupervised method for extracting multi-word units from corpora is presented and the findings examined. However, rather than evaluating the results, this article will raise the question of what constitutes a 'good' multi-word unit? The article does not claim to give any conclusive answers but perhaps instead posing a few relevant questions.