Genomic prediction for crossbred beef cattle has shown limited results using low- to moderate-density SNP panels. The relationship between the training and validation populations, as well as the size of the reference population, affects the prediction accuracy for genomic selection. Rotational crossbreeding systems require the usage of crossbred animals as sires and dams of future generations, so crossbred animals require accurate evaluation. Here, a novel method for grouping of purebred and crossbred animals (based exclusively on genotypes) for genomic selection was investigated. Clustering of animals to investigate the genetic similarity among different groups was performed using several genomic relationship criteria between individuals. Hierarchical clusters based on average-link criteria (computed as the mean distance between elements of each subcluster) were formed. The accuracy of genomic prediction was assessed using 1,500 bulls genotyped for 54,609 markers. Estimated breeding values based on all available phenotypic records for birth weight, weaning gain, postweaning gain, and yearling gain were calculated using BLUP methodologies and deregressed to ensure unbiased comparisons could be made across populations. A 5-fold validation technique was used to calculate direct genomic values for all genotyped bulls; the addition of unrelated animals in the reference population was also investigated. We demonstrate a decrease in genomic selection accuracy after including animals from disconnected clusters. A method to improve genomic selection for crossbred and purebred animals by clustering animals based on their genotype is suggested. Unlike traditional approaches for genomic selection with a fixed reference population, genomic prediction using clusters (GPC) chooses the best reference population for better accuracy of genomic prediction of crossbred and purebred animals using clustering methods based on genotypes. An overall average gain in accuracy of 1.30% was noted over all scenarios across all traits investigated when the GPC approach was implemented. Further investigation is required to assess this difference in accuracy when a larger genotyped population is available, especially for the comparison of groups with higher genetic dissimilarity, such as those found in industry-wide across-breed genetic evaluations.