Metallurgical sampling and the results of metallurgical test work form the quantitative basis for prediction of the processing behaviour of ore. These predictions are critical inputs for the design and capital and operating cost estimates for the ore processing plant, the economic evaluation of the project and the final investment decision. Ore characteristics such as hardness, competency, mineral content, grain size and texture, control ore behaviour in a mineral processing circuit. Therefore, the principles for selection of metallurgical samples should be strongly guided by orebody geology. Furthermore, only a tiny fraction of the orebody will be tested before it is mined, so metallurgical test work is always data-poor. In contrast, the geological sample database is likely to consist of tens or hundreds of thousands of records. A key principle of geometallurgy is to use the geological sample database to gain leverage from relatively few high-cost metallurgical tests. Applying geometallurgical principles to design of metallurgical sampling programmes aligns sparse test work data to abundant geological data and ensures that the data is suitable for the application of data science methods to derive robust predictions of ore processing behaviour. The selection of composite, variability, and blended samples with respect to purpose, representativity, and predictive modelling is discussed. Application of machine learning to metallurgical sample selection using multivariate data is demonstrated using case studies. The vexed question of ‘how many samples do we need?’ is addressed and a data-driven solution is proposed.