The characteristics of the signal-to-noise ratio (S/N) of Green’s functions constructed from continuous data of ambient seismic noise stacked over a 17-yr time series are shown. We used data recorded by 10 stations of the German Regional Seismic Network (GRSN). For 45 different interstation distances between 80 and 500 km, the root mean square (rms) of ambient noise in different frequency bands in the range from 0.015 to 2 Hz is used to observe seasonal as well as daily variations of noise amplitude. It is shown that rms values depend on the station location and the filter frequency. We construct the Green’s functions from noise using a window selection method (WSM) described in this paper. All data processing steps in this method are linear, avoiding nonlinear one-bit normalization. We show that the S/N of the Green’s function increases proportionally to the square root of stacking time if noise is measured at the end of the correlogram. On the other hand, S/Ns are not improved as a result of stacking the data for long time series for noise defined at the beginning of the correlogram. Therefore, stacking the Green’s function for long time series does not improve the S/N near the first arrival of the retrieved Green’s function. We interpret this observation to be the result of an inhomogeneous distribution of noise sources. The correlation functions do not perfectly converge to the Green’s functions even for stacking long time series such as a 17-yr time series.