They are ‘best’ because they maximize correlations between true values and predicted values, they are ‘linear’ because they are linear combinations of model effect estimates, and they are ‘unbiased’ because their expected value is equal to the true value.