Transfer function analysis (TFA) is a widely used method for assessing dynamic cerebral autoregulation in humans. In the present study, we assessed the test-retest reliability of established TFA metrics derived from spontaneous blood pressure oscillations and based on 5 min recordings. The TFA-based gain, phase and coherence in the low-frequency range (0.07-0.20 Hz) from 19 healthy volunteers, 37 patients with subarachnoid haemorrhage and 19 patients with sepsis were included. Reliability assessments included the smallest real difference (SRD) and the coefficient of variance for comparing consecutive 5 min recordings, temporally separated 5 min recordings and consecutive recordings with a minimal length of 10 min. In healthy volunteers, temporally separating the 5 min recordings led to a 0.38 (0.01-0.79) cm s-1 mmHg-1 higher SRD for gain (P = 0.032), and extending the duration of recordings did not affect the reliability. In subarachnoid haemorrhage, temporal separation led to a 0.85 (-0.13 to 1.93) cm s-1 mmHg-1 higher SRD (P = 0.047) and a 20 (-2 to 41)% higher coefficient of variance (P = 0.038) for gain, but neither metric was affected by extending the recording duration. In sepsis, temporal separation increased the SRD for phase by 94 (23-160)° (P = 0.006) but was unaffected by extending the recording. A recording duration of 8 min was required to achieve stable gain and normalized gain measures in healthy individuals, and even longer recordings were required in patients. In conclusion, a recording duration of 5 min appears insufficient for obtaining stable and reliable TFA metrics when based on spontaneous blood pressure oscillations, particularly in critically ill patients with subarachnoid haemorrhage and sepsis.