The method used to estimate the unfiltered longwave broadband radiance from the filtered radiances measured by the Geostationary Earth Radiation Budget (GERB) instrument is presented. This unfiltering method is used to generate the first released edition of the GERB-2 dataset. This method involves a set of regressions between the unfiltering factor (i.e.; the ratio of the unfiltered and filtered broadband radiances) and the narrowband observations of the Spinning Enhanced Visible and Infrared Imager (SEVIRI) instrument. The regressions are theoretically derived from a large database of simulated spectral radiance curves obtained by radiative transfer computations. The generation of this database is fully described. Different sources of error that may affect the GERB unfiltering have been identified and the associated error magnitudes are assessed on the database. For most of the earth–atmosphere conditions, the error introduced during the unfiltering processes is well under 0.5% (RMS error of about 0.1%). For more confidence, the unfiltered radiances of GERB-2 are validated by cross comparison with collocated and coangular Clouds and the Earth’s Radiant Energy System (CERES) observations. The agreement between the unfiltered radiances is within the science goals (1% accuracy for GERB and 0.5% for CERES) for the Flight Model 2 (FM2). For the CERES Flight Model 3 (FM3) instrument, an overall difference of 1.8% is observed. The intercomparisons indicate some scene-type dependency, which is due to the unfiltering for the cloudy scenes. This should be corrected for subsequent editions of the database.