To study fluctuations in intracardiac atrial fibrillation (AF) cycle length (CL). Sites of short AF CL may be good ablation targets, and cycle lengthening predicts ablation success. However, the optimum method for measuring AF CL, and its stability, are unclear. We hypothesized that autocorrelation better estimates AF CL than spectral dominant frequency (DF), which is susceptible to double counting, using monophasic action potentials (MAPs) to separate atrial activation from artifact. In 28 patients with paroxysmal or persistent AF, we analyzed 49 AF epochs using MAPs at the high (HRA) and low (LRA) right atrium. We estimated AF CL over 2 seconds, 10 seconds, and 2 minutes using spectral DF and autocorrelation in MAPs and filtered bipoles. In the HRA, manually measured CL was 167 +/- 25 ms. Spectral DF poorly estimated AF CL in bipolar signals (R = 0.31; P = NS), due to double counting, but accurately estimated MAP CL (R = 0.73, P < 0.001). Autocorrelation estimated MAP (R = 0.92; P < 0.001) and bipolar (R = 0.83; P < 0.001) CL, with lower errors than spectral DF (P < 0.0001). Over time, changes in DF consistently preceded reciprocal changes in organization (P < 0.001). Finally, excluding inaccurate spectra, DF and AF organization differed between HRA and LRA over 2 seconds, but correlated over 10 seconds and 2 minutes (P < 0.05). AF CL is better estimated by autocorrelation than spectral DF, particularly for bipoles, and stable when measured for >10 seconds. Notably, changes in AF CL preceded reciprocal changes in organization, yet changes in organization did not precede changes in AF CL. These results may help to interpret AF CL fluctuations.