With the current and ever-growing offering of reagents along with the vast palette of organic reactions, virtual libraries accessible to combinatorial chemists have dramatically increased in size. Yet, extracting representative subsets for experimentation is an essential step in the design of combinatorial libraries. There has been some controversy whether it is necessary to consider product properties, at some computational expense, or whether sufficiently representative sets can be identified from considerations of the reagent space alone. This study compares the efficiency of reagent-based selections and that of product-based combinatorial subsetting in the identification of representative library subsets. Quantitative estimates reported herein show that the advantage of working in product space is descriptor dependent. For some descriptors, product-based approaches provide a distinct advantage, whereas for others results from reactant pools offer comparable results. Hence the behavior of descriptors, in mapping diversity from reagent space to product space, should be investigated prior to embarking into lengthy product-based considerations. Several classes of descriptors are studied including two-dimensional fingerprints (ISIS and Daylight) and physicochemical descriptors.