Climate change projections are often given as equally weighted averages across ensembles of climate models, despite the fact that the sampling of the underlying ensembles is unclear. We show that a hierarchical clustering of a metric of spatial and temporal variations of either surface temperature or precipitation in control simulations can capture many model relationships across different ensembles. Strong similarities are seen between models developed at the same institution, between models sharing versions of the same atmospheric component, and between successive versions of the same model. A perturbed parameter ensemble of a model appears separate from other structurally different models. The results provide insight into intermodel relationships, into how models evolve through successive generations, and suggest that assuming model independence in such ensembles of opportunity is not justified.