According to him, any typological approach involving studying/categorizing any behaviors other than these private/personal cognitive tendencies constitutes "folk typology" and is therefore almost entirely worthless.
He justifies this by saying that people will often behave in ways contrary to their type due to external pressures, which is true, but you can still observe behavioral trends and make inferences about people's types that become increasingly accurate as you gather more information.
The fact is that each type really does tend toward certain behavioral patterns on average, so you can use induction to make reasonably good guesses--but this is unacceptable to SW because it doesn't focus exclusively on "cognitive tendencies" and requires induction/guesswork. (Basically, it requires induction via extroverted perception and doesn't fit squarely into his little Ti+Si box, so it's worthless.)
SW is kind of vague in explaining which behaviors qualify as relevant to typology--he says only a person's most private thoughts and writings about the nature of life can be used as legitimate typological data, and that using any other sort of behavior is inherently invalid and a total waste of time.