The GATE software platform, based on the Geant4 toolkit for simulating particle interactions with matter, enables simulation of, among other medical imaging and treatment systems, positron emission tomography. However, at least one publication (Moraes et al 2015 Phys. Med. 31 43-8) has reported discrepancies between the expected results and those obtained using GATE simulations, specifically with respect to the coincidence sorter which processes single events detected by the scanner to find coincidence pairs. In particular, the current software appears to overestimate the number of 'true' coincidence pairs when in multi-window mode, while the delayed coincidence window, used to estimate the randoms present in the prompt coincidence window, underestimates the randoms. Both effects are particularly evident at high count rates. We have investigated this discrepancy and reproduced the reported problems. We have also rewritten the relevant portion of the GATE code to correct the issue. In this note we describe the modifications to the coincidence sorter and repeat the simulations which previously showed unexpected results. Some discrepancies remain in the estimation of the randoms with the single-window mode which are a consequence of the algorithm itself. In multi-window mode however, the simulation agrees exactly with the expected results. The modifications to the coincidence sorter code will be incorporated into the next release of GATE (> version 7.2).