The effect of introducing the universal quantifier ‘all’ into the class inclusion question was investigated using 104 children aged 59 to 90 mo. One group of children was asked the standard version of the question, another an ‘all-subset’ version in which ‘all’ preceded the subclass, the third an ‘all-superset’ version with ‘all’ before the superordinate class, and the fourth a ‘double-all’ version with ‘all’ in both locations. When the superordinate class was mentioned last in the question, performance was better on the all-superset and double-all versions than on the standard version. When the subclass was mentioned last, performance was better on the all-superset version only. Performance on the all-subset version did not differ from that on the standard version. The results were explained in terms of the attention-directing role of ‘all’, together with the proposal that performance improves if attention is drawn towards the superordinate class and/or away from the contrast between the subclasses.