The aim of this paper is to confront some logically and philosophically puzzling problems, raised by philosophers and cognitive psychologists, with a semantic description of the concept of 'similarity' and with the meaning and the use of operators of similarity in certain languages. 'Similarity' is shown to be a cluster of notions rather than a unitary concept. By contrast, there is a certain unity of the items in the group of similarity subtypes, in that a unifying mechanism underlies them and justifies gathering them under one general, although vague concept. The various subtypes of similarity (such as duplication, type/token relations, analogy, etc.) show various modes of equilibrium between unifying and separating elements. A study of the nature and the etymology of grammatical and lexical operators of similarity (e.g. 'as', 'like', '(the) same' in English; 'kemo', 'oto', 'dome' in Hebrew; and 'als' in German) supports this claim. Finally, as to Austin's claim that similarity words are adjusters, the present study exemplifies this by showing how similarity operators add to the basic discrete nature of the processes of labeling, recognizing and naming, a continuous counterpart which affords flexibility and expandability to the cognitive process. The puzzling nature of the concept of similarity is shown to be rooted in this peculiar cognitive function.