We compare the ability of various site‐condition proxies (SCPs) to reduce the aleatory variability of ground motion prediction equations (GMPEs). Three SCPs (measured V _{S 30} , inferred V _{S 30} , local topographic slope) and two accelerometric databases (RESORCE and NGA‐West2) are considered. An artificial neural network (ANN) approach including a random‐effect procedure is used to derive GMPEs setting the relationship between peak ground acceleration ( PGA ), peak ground velocity ( PGV ), pseudo‐spectral acceleration [ PSA ( T )], and explanatory variables ( M _w , R _JB , and V _{S 30} or Slope ). The analysis is performed using both discrete site classes and continuous proxy values. All “non‐measured” SCPs exhibit a rather poor performance in reducing aleatory variability, compared to the better performance of measured V _{S 30} . A new, fully data‐driven GMPE based on the NGA‐West2 is then derived, with an aleatory variability value depending on the quality of the SCP. It proves very consistent with previous GMPEs built on the same data set. Measuring V _{S 30} allows for benefit from an aleatory variability reduction up to 15%.