Models to predict herbage intake were constructed using 168 dairy cow records from three grazing experiments. Variables included fell into three categories: animal state, sward state and animal behaviour. Linear regression models of varying complexity were obtained by removing variables from the best fitting model to reflect progressive lack of information availability on farms. Thus, behavioural variables were removed first, followed by sward surface height and milk fat concentration. Models were subject to outlier analysis and collinearity tests. Equivalent models were constructed using ridge regression to minimize collinearity problems. They were tested using 20 Holstein–Friesian dairy cows continuously stocked on a perennial ryegrass sward. A `best practice' treatment [7 cm sward surface height (SSH), 6 kg day−1 concentrate (C)] was used together with treatments of SSH5/C6, SSH7/C8, SSH7/C0 and SSH9/C6. The best model accounted for 0.37 of the variance in the estimation data and contained the following variables: concentrate intake, milk yield, milk fat concentration, days in milk, sward surface height and chewing rate while ruminating. Model performance against test data was generally poor. This was mainly because of consistent underprediction of herbage intake, caused in part by the higher average herbage intakes in the test data compared with the estimation data.