Unmodeled displacements in GNSS times series, induced by instrumental artifacts or geophysical events, create significant biases in station trajectory parameters that can propagate into the reference frame itself. While non-tectonic ‘jumps’, such as equipment changes, affect only a specific GNSS station, seismically-induced displacements can affect large numbers of sites, severely threatening the frame’s stability. Manually reviewing individual GNSS time series for such effects is highly impractical because there can be thousands of GNSS stations in a frame, and the total number of earthquakes Mw ≥ 6.0 since GPS became fully operational is + 5100. To avoid this time-consuming task, automated methods rely on empirical power-law functions to determine which earthquake-station pairs require coseismic displacement parameters. Still, ‘conservative’ power-law functions tend to add coseismic offsets to stations that do not need them, which can also threaten the stability of the frame. In this work, we present an empirical formulation that was obtained using 809 global seismic events to fit power-law parameters that do not overestimate the region of influence of earthquakes. Our method is based on a two-level selection process: level 1 is isotropic and only considers the epicentral distance between the stations and the earthquake, and level 2 uses the geophysical parameters of the earthquake to predict a ‘tighter’ displacement pattern to select which stations require coseismic trajectory parameters. We applied our level 2 method to a database of ~ 4700 event-station pairs and showed that it removed ~ 55% of the total pairs, all of which had been falsely selected by level 1.