Microarray technology allows gene expression profiling at a global level. Many algorithms for the normalization of raw microarray data have been proposed, but no attempt has yet been made to propose operationally verifiable criteria for their comparative evaluation, which is necessary for the selection of the most appropriate method for a given dataset. This study develops a set of operational criteria for assessing the impact of various normalization algorithms in terms of accuracy (bias), precision (variance) and over-fitting (information reduction). The use of these criteria is illustrated by applying the three most widely used algorithms (global median normalization, spiked-in based normalization and lowess) on a specifically designed, multiply-controlled dataset.