It is well known that the first DNA polymerase, Taq, is quite error prone. Newer generation commercial enzymes that have either been isolated from different thermophile species or have been improved by recombination are less error prone. How are these error rates compared? For instance, if this is done by Sanger sequencing, the average signal will dominate in reading the output and so it will be very unlikely to pick up on errors by this method.
According to their website New England Biolabs use a version of the approach pioneered by Wayne Barnes, as described in:
Kermekchiev, M.B., Tzekov, A and Barnes, W.M. (2003) Nucl. Acids Res. 31, 6139–6147
This is basically an assay for the mutation rate in a PCR-amplified lacZ (β-galactosidase) gene, assayed by transforming E. coli, plating on the chromogenic β-galactosidase substrate Xgal, and then scoring white colonies as mutated genes. Also according to NEB, Agilent Technologies use a similar mutational assay, but based upon the lacI (lac repressor) gene.