A growing empirical literature is analyzing the effects of weather fluctuations on a variety of economic outcomes with the goal of better understanding the potential impacts of climate change. In agricultural studies, constructing weather variables typically requires researchers to define a “season”, a time period over which weather conditions are considered relevant to the agricultural outcome of interest. While researchers often have the background knowledge to make reasonable assumptions about seasonality in crop-specific analyses, these modeling choices are less obvious when dealing with aggregate agricultural data encompassing multiple crops or livestock. In this article, we explore the consequences of assuming an incorrect season in such analyses. We first provide a conceptual framework to show that imposing an incorrect season essentially introduces non-classical measurement error in weather regressors, causing unknown biases in weather impacts. We confirm this finding in simulations. We then propose a tractable data-driven approach to recover the “true” underlying season. The approach consists of a grid search with cross-validation that evaluates the fit of models based on a wide range of season definitions. In simulations, we find the approach is effective at recovering the “true” season under certain data generating processes. Finally, we apply our approach to a US state-level panel of agricultural Total Factor Productivity. We find, unsurprisingly, considerable differences in seasonality across regions. Importantly, our empirical findings suggest that imposing arbitrary seasons lead to substantially different estimates of weather effects in either direction, in line with our theoretical and simulated results. This work contributes to the development of more robust empirical studies of climate change impacts on agriculture and beyond.