We consider clustering in group decision making where the opinions are given by pairwise comparison matrices. In particular, the k-medoids model is suggested to classify the matrices as it has a linear programming problem formulation. Its objective function depends on the measure of dissimilarity between the matrices but not on the weights derived from them. With one cluster, our methodology provides an alternative to the conventional aggregation procedures. It can also be used to quantify the reliability of the aggregation. The proposed theoretical framework is applied to a large-scale experimental dataset, on which it is able to automatically detect some mistakes made by the decision-makers.