Upon a Power on Reset event the expected performance is not very accurate as it adopts an OCV method to estimate the SoC before the coulomb-counter improves the accuracy, but this takes time. The expected error convergence is approximately half the error in half an hour. This means that if the initial estimate missed the mark by 30% then in half hour it would converge to 15% and so on.
It is not recommend to disconnect the power from the fuel gauge at all but instead make use of the low power operating modes, the M5 algorithm is most accurate after a few discharge cycles it is not expected to continuously remove the battery. It will be able to learn but this will take time after each POR event.