The great majority of data processing methods for magnetotelluric measurements are based on an average Fourier spectral analysis to estimate either the transfer function or the coherence function. This assumes that the time‐series data are stationary. The aim of this paper is to present a preprocessing method able to extract the stationary and coherent part of the original signals. The practitioner can then apply the magnetotelluric method of his choice to these new data. This preliminary data sorting is done in four steps: (i) slow drifts are eliminated using a high‐pass filter whose cut‐off frequency is determined by an iterative procedure. Based on run tests, this procedure is also able to remove segments with non‐independent samples in the time series; (ii) non‐stationary segments are eliminated after band‐pass filtering; (iii) non‐coherent segments are eliminated before spectral analysis; (iv) the impedance tensor value is then retained, at a given frequency, only if the signals are coherent. This preprocessing method was tested on the simplest, but still used, magnetotelluric method which uses only two field components, and it was found that the average resistivity standard deviation decreased significantly from 14.6 Ωm without sorting to 8.6 Ωm after sorting.