This paper provides a corpus-driven investigation into establishing classes of nouns based on grammatical environments relevant to countability, such as combination with cardinal modifiers or appearing as a bare singular. We investigate the countability environments of Allan (1980) and assess their predictive power across a large corpus (350 million words). We show, by applying machine learning methods, that while the environments Allan (1980) distinguishes are predictive, the occurrence of nouns as bare singular and/or bare plural is substantially more powerful as a diagnostic. Using the most important environments, we induce, through automatic clustering, a set of countability classses, which distinguish between varieties of countable, non-countable and pluralia tantum nouns.