Your observation are quite common..
Unless you hack into the crystal circuit and alter the loading capacitance to tune the accuracy .
Tuning the crystal is also hit and miss as the loading of a freq counter can alter the circuit conditions.
For Most RTC's they are are in the +/- 2-3 mins/month..


An excellent alternative is the ST M41T00 RTC that is I2c with an initial accuracy of +/- 1 mins/month ,but the remaining inaccuracy can be dialled out in software ,say by the end user in a simple calibration routine..