We lack a comprehensive understanding of evolutionary pattern and process because short-term and long-term data have rarely been combined into a single analytical framework. Here we test alternative models of phenotypic evolution using a dataset of unprecedented size and temporal span (nearly 8,000 data points). The data are body-size measurements taken from historical studies, the fossil record, and among-species comparative data representing mammals, squamates, and birds. By analyzing this unusually large dataset, we identify stochastic models that can explain evolutionary patterns on both short and long timescales and reveal a remarkably consistent pattern in the timing of divergence across taxonomic groups. Even though rapid, short-term evolution often occurs in intervals shorter than 1 Myr, the changes are constrained and do not accumulate over time. Over longer intervals (1–360 Myr), this pattern of bounded evolution yields to a pattern of increasing divergence with time. The bestfitting model to explain this pattern is a model that combines rare but substantial bursts of phenotypic change with bounded fluctuations on shorter timescales. We suggest that these rare bursts reflect permanent changes in adaptive zones, whereas the short-term fluctuations represent local variations in niche optima due to restricted environmental variation within a stable adaptive zone.