The purpose of this study is to characterize and understand the long-term behavior of the output from megavoltage radiotherapy linear accelerators. Output trends of nine beams from three linear accelerators over a period of more than three years are reported and analyzed. Output, taken during daily warm-up, forms the basis of this study. The output is measured using devices having ion chambers. These are not calibrated by accredited dosimetry laboratory, but are baseline-compared against monthly output which is measured using calibrated ion chambers. We consider the output from the daily check devices as it is, and sometimes normalized it by the actual output measured during the monthly calibration of the linacs. The data show noisy quasi-periodic behavior. The output variation, if normalized by monthly measured "real' output, is bounded between ± 3%. Beams of different energies from the same linac are correlated with a correlation coefficient as high as 0.97, for one particular linac, and as low as 0.44 for another. These maximum and minimum correlations drop to 0.78 and 0.25 when daily output is normalized by the monthly measurements. These results suggest that the origin of these correlations is both the linacs and the daily output check devices. Beams from different linacs, independent of their energies, have lower correlation coefficient, with a maximum of about 0.50 and a minimum of almost zero. The maximum correlation drops to almost zero if the output is normalized by the monthly measured output. Some scatter plots of pairs of beam output from the same linac show band-like structures. These structures are blurred when the output is normalized by the monthly calibrated output. Fourier decomposition of the quasi-periodic output is consistent with a 1/f power law. The output variation appears to come from a distorted normal distribution with a mean of slightly greater than unity. The quasi-periodic behavior is manifested in the seasonally averaged output, showing annual variability with negative variations in the winter and positive in the summer. This trend is weakened when the daily output is normalized by the monthly calibrated output, indicating that the variation of the periodic component may be intrinsic to both the linacs and the daily measurement devices. Actual linac output was measured monthly. It needs to be adjusted once every three to six months for our tolerance and action levels. If these adjustments are artificially removed, then there is an increase in output of about 2%-4% per year.