It has become an article of faith among many functional and cognitive linguists that the complex abstract structures posited by generative grammarians are an artifact of 'disembodied sentences that analysts have made up ad hoc, rather than utterances produced by real people in real discourse situations' (Michael Tomasello). Their view is that if one focuses on 'naturally occurring discourse', then grammar will reveal itself to be primarily a matter of memorized formulas and simple constructions. This paper challenges that view. Basing its claims on a 170MB corpus of conversational English, it argues that the nature of real discourse reinforces the need for a sophisticated engine for representing grammatical knowledge.