Predicting protein fractions and coagulation properties in bovine milk using Fourier transform infrared (FT-IR) measurements is desirable. However, such predictions may rely on correlations with total protein content. The aim of this study was to show how correlations between total protein content, protein fractions, and coagulation properties are responsible for the successful prediction of protein fractions and rennet-induced coagulation properties in milk samples. This study comprised 832 bovine milk samples from 2 breeds (426 Holstein and 406 Jersey). Holstein samples were collected from 20 Danish dairy herds from October to December 2009; Jersey samples were collected from 22 Danish dairy herds from February to April 2010. All samples were from conventional herds and taken while cows were housed. The results showed that κ-CN, αS1-CN, αS1-CN with 8 phosphorylated groups attached (αS1-CN 8P), and curd firming rate could be predicted from FT-IR measurements of the milk samples (with coefficients of determination between 0.66 and 0.71). However, the success of these FT-IR-based predictions was based on indirect relationships with total protein content. Hence, the FT-IR predictions relied on covariance structures with total protein content rather than absorption bands directly associated with the protein fractions and coagulation properties. If covariance structures between the protein fractions, coagulation properties, and total protein content used to calibrate partial least squares models were not conserved in future samples, these samples would show incorrect predictions of the protein fractions and coagulation properties. We demonstrated this using samples from 1 breed to calibrate and samples from the other breed to validate partial least squares models for β-CN. The 2 breeds had different covariance structures between β-CN and total protein content, and the validation samples yielded incorrect predictions. This finding may limit the usefulness of FT-IR-based predictions of protein fractions in milk recording, because indirect covariance structures in the calibration set must be valid for future samples, or future samples will show incorrect predictions.