Surface-enhanced Raman scattering (SERS) has become a widely used spectroscopic technique for chemical identification, providing unbeaten sensitivity down to the singlemolecule level. The amplification of the optical near field produced by collective electron excitations plasmons in nanostructured metal surfaces gives rise to a dramatic increase by many orders of magnitude in the Raman scattering intensities from neighboring molecules. This effect strongly depends on the detailed geometry and composition of the plasmonsupporting metallic structures. However, the search for optimized SERS substrates has largely relied on empirical data, due in part to the complexity of the structures, whose simulation becomes prohibitively demanding. In this work, we use state-of-the-art electromagnetic computation techniques to produce predictive simulations for a wide range of nanoparticle-based SERS substrates, including realistic configurations consisting of random arrangements of hundreds of nanoparticles with various morphologies. This allows us to derive rules of thumb for the influence of particle anisotropy and substrate coverage on the obtained SERS enhancement and optimum spectral ranges of operation. Our results provide a solid background to understand and design optimized SERS substrates.