We address the bioinformatic issue of accurately separating amplified genes of the major histocompatibility complex (MHC) from artefacts generated during high-throughput sequencing workflows. We fit observed ultra-deep sequencing depths (hundreds to thousands of sequences per amplicon) of allelic variants to expectations from genetic models of copy number variation (CNV). We provide a simple, accurate and repeatable method for genotyping multigene families, evaluating our method via analyses of 209 b of MHC class IIb exon 2 in guppies (Poecilia reticulata). Genotype repeatability for resequenced individuals (N = 49) was high (100%) within the same sequencing run. However, repeatability dropped to 83.7% between independent runs, either because of lower mean amplicon sequencing depth in the initial run or random PCR effects. This highlights the importance of fully independent replicates. Significant improvements in genotyping accuracy were made by greatly reducing type I genotyping error (i.e. accepting an artefact as a true allele), which may occur when using low-depth allele validation thresholds used by previous methods. Only a small amount (4.9%) of type II error (i.e. rejecting a genuine allele as an artefact) was detected through fully independent sequencing runs. We observed 1–6 alleles per individual, and evidence of sharing of alleles across loci. Variation in the total number of MHC class II loci among individuals, both among and within populations was also observed, and some genotypes appeared to be partially hemizygous; total allelic dosage added up to an odd number of allelic copies. Collectively, observations provide evidence of MHC CNV and its complex basis in natural populations.