Semantic memory retrieval was investigated in 2 experiments. In the first experiment, 5s were shown a category name and asked to respond with a word belonging to the category (for one block of trials) or a class to which the category name belonged (for another block of trials). Subjects produced a category member faster than they produced a superordinate. For example, they produced an instance of the category CAR faster than they produced a superordinate such as vehicle. The time taken to retrieve a superordinate was strongly related to the category's hierarchical position, while the time taken to retrieve an instance was not so related. In the second experiment, 5s produced free associates to categories differing in hierarchical level. More subordinates were given for all but the lowest level categories. The data argue against the notion that the superset is the most accessible property of a category or concept.