Taxonomic classification of the thousands-millions of 16S rRNA gene sequences generated in microbiome studies is often achieved using a naïve Bayesian classifier (for example, the Ribosomal Database Project II (RDP) classifier), due to favorable trade-offs among automation, speed and accuracy. The resulting classification depends on the reference sequences and taxonomic hierarchy used to train the model; although the influence of primer sets and classification algorithms have been explored in detail, the influence of training set has not been characterized. We compared classification results obtained using three different publicly available databases as training sets, applied to five different bacterial 16S rRNA gene pyrosequencing data sets generated (from human body, mouse gut, python gut, soil and anaerobic digester samples). We observed numerous advantages to using the largest, most diverse training set available, that we constructed from the Greengenes (GG) bacterial/archaeal 16S rRNA gene sequence database and the latest GG taxonomy. Phylogenetic clusters of previously unclassified experimental sequences were identified with notable improvements (for example, 50% reduction in reads unclassified at the phylum level in mouse gut, soil and anaerobic digester samples), especially for phylotypes belonging to specific phyla (Tenericutes, Chloroflexi, Synergistetes and Candidate phyla TM6, TM7). Trimming the reference sequences to the primer region resulted in systematic improvements in classification depth, and greatest gains at higher confidence thresholds. Phylotypes unclassified at the genus level represented a greater proportion of the total community variation than classified operational taxonomic units in mouse gut and anaerobic digester samples, underscoring the need for greater diversity in existing reference databases.