ABSTRACT Marchigiana, Chianina, and Romagnola are three Italian autochthonous beef cattle breeds that have been historically selected for meat production. Recent advancements suggest that the use of genomic data and multi‐breed (MB) models to combine information from different breeds may help to increase the accuracies of genomic predictions, in particular if the available data per breed is limited. This study aimed to evaluate and compare the accuracies of genomic predictions for average daily gain (ADG) in the three Italian breeds. We implemented different scenarios using phenotypes collected on 5303 young bulls in performance tests across the three breeds, 23,793 pedigree records, and 4593 genotypes, and then validated through the linear regression method. The implemented scenarios were: pedigree Best Linear Unbiased Prediction (pBLUP) and single‐step Genomic BLUP (ssGBLUP) single‐trait single‐breed evaluations where each breed was modelled separately; pBLUP and ssGBLUP single‐trait multi‐breed evaluations where ADG was modelled as the same trait for all breeds, and ssGBLUP multi‐trait multi‐breed evaluations where ADG was considered as a different correlated trait across breeds. In addition, single‐ and multi‐breed pBLUP and ssGBLUP evaluations were implemented including weight at 1 year of age and muscularity as correlated traits of ADG in a multi‐trait approach. Results highlighted the improved accuracies (an average of 5% in ssGBLUP models compared to corresponding pBLUP ones) when incorporating genomic data in the prediction models. Moreover, single‐trait multi‐breed scenarios resulted in higher accuracy for breeds with lower heritabilities for ADG (an average of 4% for single‐trait multi‐breed models compared to single‐breed ones), confirming the importance of leveraging data from populations with higher heritabilities. Lastly, adding two correlated traits next to ADG in the single‐ and multi‐breed ssGBLUP yielded even higher accuracies than the scenarios only encompassing ADG. The observed increases in accuracy when leveraging data from more populations and/or more traits could be helpful when implementing genomic predictions for innovative traits with limited records per individual or low heritabilities, and for the genetic improvement of local populations where limited data availability represents a challenge for traditional genetic selection.