To screen heterogeneous catalysts in silico, the linear energy relationships derived from the Brønsted−Evans−Polanyi principle are extremely useful. They connect the reaction energy of a given elementary step to its activation energy, hence providing data that can be fed to kinetics models at a minimal cost. However, to ensure reasonable predictions, it is essential to control the statistical error intrinsic to this approach. We derived several types of linear energy relations for a series of CH and OH bond scissions in simple alcohol molecules on compact facets of seven transition metals (Co, Ni, Ru, Rh, Pd, Ir, and Pt) aiming at a single but accurate relation. The quality of the relation depends on its nature and/or on the manner the data are split: a single linear relation can be constructed for all metals together on the basis of the original Brønsted−Evans−Polanyi formulation with a mean absolute error smaller than 0.1 eV, whereas the more recent transition state scaling approach requires considering each metal individually to reach an equivalent accuracy. In addition, a close statistical analysis demonstrates that errors stemming from such predictive models are not uniform along the set of metals and of chemical reactions that is considered opening the road to a better control of error propagation.