Referencing is frequently used to remove common-mode signals from multielectrode data, in both freely moving animals and in vitro preparations. For action potential (AP) detection, referencing by subtracting the common average signal has been shown to increase AP signal-to-noise ratio (SNR). This method fails, however, when large transients occur on individual electrodes, as occurs during electrical stimulation or with large APs during spontaneous recordings. To deal with these cases, we propose using the common median as a reference. The common median has an improved SNR for AP detection (leading to more isolated single units and more detected APs per unit) and, unlike common average referencing, does not generate spurious APs when processing large single-electrode transients.