Think-aloud protocols are a dominant method in usability testing. There is, however, only little empirical evidence on the actual validity of the method. This paper describes an experiment that compares concurrent and retrospective think-aloud protocols for a usability test of an online library catalogue. There were three points of comparison: usability problems detected, overall task performance, and participant experiences. Results show that concurrent and retrospective think-aloud protocols reveal comparable sets of usability problems, but that these problems come to light in different ways. In retrospective think-aloud protocols, more problems were detected by means of verbalisation, while in concurrent think-aloud protocols, more problems were detected by means of observation. Moreover, in the concurrent think-aloud protocols, the requirement to think aloud while working had a negative effect on the task performance. This raises questions about the reactivity of concurrent think-aloud protocols, especially in the case of high task complexity.