In a study over a number of months, involving over 500 milk producers in three locations, and representing bulk milk collection from cans as well as mobile and stationary refrigerated bulk milk tanks, the precision of various random sampling and testing frequencies in estimating the fat, protein and lactose content of milk was evaluated. In addition, the traditional procedure of estimating the composition of milk supplies from chemically preserved composites was compared to values obtained by analysis of fresh samples. While sampling and testing every milk collection gave the most precise estimate of milk composition for each producer, it was found that for producers using refrigerated bulk milk tanks with ex‐farm collection, four random samples per month gave a precision of ±4% (or ±0.14% fat approx) for monthly fat content and ±l.2% (or ±0.04% fat approx) for annual fat content. Due to lower variability the precision for both protein and lactose estimation was much higher. The sample compositing procedure, while capable of a relatively high level of precision, did not always achieve this in practice and tended to underestimate the constituents in milk, especially protein and la