It maps the linear predictor $\eta = X\beta$ to the mean of the distribution $\mu$.

It transforms the response variable to make it normal.

It defines the variance structure of the errors.

It ensures the errors are independent.