Core Ideas Efficiency in genomic selection is not particularly based on detailed genotype profiling facilitated by maximum marker density. Extensive genome‐wide linkage disequilibrium is a common characteristic of breeding pools in many crop species. Every quantitative trait locus across the genome can be captured by one or a few representative markers. Fewer representative markers selected in respect of linkage disequilibrium (LD) can capture the association between a genomic region and a phenotypic trait. Low‐density marker sets enable genomic prediction accuracies in breeding populations with strong LD comparable to those achieved with high‐density genotyping. Genomic selection (GS) has revolutionized breeding for quantitative traits in plants, offering potential to optimize resource allocation in breeding programs and increase genetic gain per unit of time. Modern high‐density single nucleotide polymorphism (SNP) arrays comprising up to several hundred thousand markers provide a user‐friendly technology to characterize the genetic constitution of whole populations and for implementing GS in breeding programs. However, GS does not build upon detailed genotype profiling facilitated by maximum marker density. With extensive genome‐wide linkage disequilibrium (LD) being a common characteristic of breeding pools, fewer representative markers from available high‐density genotyping platforms could be sufficient to capture the association between a genomic region and a phenotypic trait. To examine the effects of reduced marker density on genomic prediction accuracy, we collected data on three traits across 2 yr in a panel of 203 homozygous Chinese semiwinter rapeseed (Brassica napus L.) inbred lines, broadly encompassing allelic variability in the Asian B. napus genepool. We investigated two approaches to selecting subsets of markers: a trait‐dependent strategy based on genome‐wide association study (GWAS) significance thresholds and a trait‐independent method to detect representative tag SNPs. Prediction accuracies were evaluated using cross‐validation with ridge‐regression best linear unbiased predictions (rrBLUP). With semiwinter rapeseed as a model species, we demonstrate that low‐density marker sets comprising a few hundred to a few thousand markers enable high prediction accuracies in breeding populations with strong LD comparable to those achieved with high‐density arrays. Our results are valuable for facilitating routine application of cost‐efficient GS in breeding programs.