Genomics offers new opportunities for the effective selection of novel traits. For traits such as mastitis resistance, hoof health, or milk composition records based on medium infrared (MIR) data, for example, enough records are usually available to carry out genomic evaluations based on sire genotypes and the phenotypes of their daughters. For traits that are more novel or expensive to collect, such as individual feed intake or immune response, the development of a cow reference population is the most effective approach. The reliability of the resulting genomic predictions depends primarily on the size of the reference population and on trait heritability, as shown by Daetwyler et al. (2008). To provide an empirical check of these theoretical estimates of reliability, the reliability of genomic selection was estimated for various traits using a reference population of 1,000 to 10,000 Canadian-born Holstein cows that had been genotyped with a panel of 6,000 SNP or more. All genotypes were imputed to 50K. The effects of SNP were estimated from cow records only, after excluding the dams of validation bulls. Bulls first proven in 2013 and 2014 were then used to carry out a validation and estimate the accuracy of genomic selection based on these SNP effects. Differences between accuracies obtained this way and using the Daetwyler formula are reported for traits of varying heritability and degree of indirect selection. Results confirm that large reference populations are usually required to achieve adequate accuracy. In many instances, the accuracy of genomic selection for novel traits can be increased through the use of indicator traits. Cow size and MIR data are used as examples to show how they can increase the accuracy of genomic selection for feed efficiency. Expected rates of genetic progress are calculated for each scenario, using the selection intensities and generation intervals currently realized in North American dairy cattle.