One of the steps in a kansei engineering study is the evaluation of prototypes by users. Usually, the participants in the study give ratings to each one of the prototypes either using a semantic differential method or a Likert scale. Normally, the number of prototypes that can be rated is limited due to the following two reasons: (1) cost (the more prototypes to be prepared, even if they are simple mock-ups, the more expensive the study will be), (2) time needed for the evaluation (having too many prototypes will make the rating too time consuming, with the added problem of tiredness of participants). Although not having too many prototypes in the study, too often a lot of properties (factors) have to be studied. This combination of many factors with few prototypes creates a challenging situation when we want to link our responses (the kansei words) with our product properties using some kind of regression analysis. Although this problem is often marginalised and not explicitly considered in the literature describing kansei engineering studies, how the prototypes for the study are selected is an issue of great importance to achieve good quality results. In order to obtain sets of prototypes adequate for the later use of statistical regression models, many studies suggest selecting the prototypes following a full factorial design, or a Taguchi orthogonal array. Of course, when prototypes are created for the study, it is possible to make them in a way that they properly define a matrix with the needed characteristics (although using a full factorial design often leads to too many prototypes, so not all of them can be produced). However, using a selection of already existing products as prototypes is very common, and the degrees of freedom in this situation are deeply reduced. The idea then is finding existing products that follow the needed combinations of properties: this is not always easy to achieve, and sometimes it is impossible. The purpose of this paper is twofold. First, analyse the problems that arise when not using a well-though selection of prototypes (an issue too often neglected), based on a series of examples. Second, give recommendations on how to select the minimum best set of prototypes that allow a correct interpretation of results.