Best practices are currently being developed for the acquisition and processing of resting-state magnetic resonance imaging data used to estimate brain functional organization—or “functional connectivity.” Standards have been proposed based on test–retest reliability, but open questions remain. These include how amount of data per subject influences whole-brain reliability, the influence of increasing runs versus sessions, the spatial distribution of reliability, the reliability of multivariate methods, and, crucially, how reliability maps onto prediction of behavior. We collected a dataset of 12 extensively sampled individuals (144 min data each across 2 identically configured scanners) to assess test–retest reliability of whole-brain connectivity within the generalizability theory framework. We used Human Connectome Project data to replicate these analyses and relate reliability to behavioral prediction. Overall, the historical 5-min scan produced poor reliability averaged across connections. Increasing the number of sessions was more beneficial than increasing runs. Reliability was lowest for subcortical connections and highest for within-network cortical connections. Multivariate reliability was greater than univariate. Finally, reliability could not be used to improve prediction; these findings are among the first to underscore this distinction for functional connectivity. A comprehensive understanding of test–retest reliability, including its limitations, supports the development of best practices in the field.