The influence of levels of abstraction in picture-word matching was examined. The items each consisted of one picture and three successively presented words. Hierarchies with words for superordinate, basic, and subordinate level concepts were used (e.g., plant, flower, rose). The picture-word condition (congruent, incongruent), the word position (first, second, third), and the level of categorization (subordinate, basic, superordinate) were manipulated. Reaction times, error rates, and pupillary responses were recorded. Pupillary responses coincided with behavioral data. In general, there was an advantage for subordinate and basic level processing compared to superordinate level processing. However, switches to words for superordinate concepts were most facilitated. These findings support a two-step account of picture-word matching. First, the picture is categorized according to its concrete features. Second, amodal features are processed.