Customer segmentation is usually the first step towards customer analysis and helps to make strategic plans for a company. Similarity between customers plays a key role in customer segmentation, and is usually evaluated by distance measures. While various distance measures have been proposed in data mining literature, the desirable distance measures for various data sources and given application domains are rarely known. One of the reasons lies in that semantic meaning of similarity and distance measures is usually ignored. This paper discusses several issues related to evaluating customer similarity based on their transaction data. Various set distance measures for customer segmentation are analyzed in several imaginary scenarios, and it is shown that each measure has different characteristics which make the measure useful for some application domains but not for others. We argue that no measure always performs better than other measures, and suitable measures should be adopted for specific purposes depending on applications.