Abstract Key message Based on experimental and simulated data for maritime pine ( Pinus pinaster Ait.) in a genomic selection context, our study reveals that the often-highlighted equivalence between genome-based and pedigree-based prediction accuracies of breeding values is associated with a zero accuracy of genome-based prediction within families, which can be attributed to the still insufficient size of the genomic training sets for conifers. Context Genomic selection is a promising approach for forest tree breeding. However, its advantage in terms of prediction accuracy over conventional pedigree-based methods is unclear and within-family accuracy is rarely assessed. Aims We used a pedigree-based model (ABLUP) with corrected pedigree data as a baseline reference for assessing the prediction accuracy of genome-based model (GBLUP) at the global and within-family levels in maritime pine ( Pinus pinaster Ait). Methods We considered 39 full-sib families, each comprising 10 to 40 individuals, to constitute an experimental population of 833 individuals. A stochastic simulation model was also developed to explore other scenarios of heritability, training set size, and marker density. Results Prediction accuracies with GBLUP and ABLUP were similar, and within-family accuracy with GBLUP was on average zero with large variation between families. Simulations revealed that the number of individuals in the training set was the principal factor limiting GBLUP accuracy in our study and likely in many forest tree breeding programmes. Accurate within-family prediction is possible if 40–65 individuals per full-sib family are included in the genomic training set, from a total of 1600–2000 individuals in the training set. Conclusions The increase in the number of individuals per family in the training set lead to a significant advantage of GBLUP over ABLUP in terms of prediction accuracy and more clearly justify the switch to genome-based prediction and selection in forest trees.Read less <