Tag clouds are popular features on web pages, not only to support browsing but also to provide an overview over the content of the page or to summarize search retrieval results. Commonly, the arrangement of tags is based on a random layout or an alphabetic ordering of the tags. Previous research suggests to further structure the tag clouds according to semantics, typically employing cooccurrence-based relations to assess the semantic relatedness of two tags. Regarding the layout of the resulting structure, a wide variety of representations has been proposed. However, only few papers motivate their design choice or evaluate its performance from the perspective of a user, leaving it open if the approach answers the users' expectations. In this paper we present the results of a study in which we observed how humans structure user-generated tags of a social bookmarking system given the task that the resulting layout should provide a quick overview over a search retrieval result. We examine the participants' layouts based on the final arrangement of tags and a detailed interview conducted after the task. Thereby, we analyze and characterize the different term relations employed as well as the higher-level structures generated. The deeper understanding of what criteria are considered important by humans can inform the design of automatic algorithms as well as future studies evaluating their performance.