Predicting herbivore control over plants (i.e. changes in plant mass due to herbivore damage) is a central goal of ecology. Progress has been limited, however, because the vegetation characteristics thought to influence herbivore control are naturally correlated and typically experimentally confounded. To address this problem, we defined eight conventional models that predict herbivore control over plant community mass, each model based on a different vegetation characteristic (i.e. host concentration, tissue nitrogen, growth rate, size, tolerance of herbivory or net primary productivity). We then used structural equation modelling to test each model against two field experiments. Our results clearly rejected all models except for a tolerance of herbivory mechanism; stems with greater access to limiting resources better tolerated herbivory, regardless of where herbivore activity was greatest. Consequently, herbivore reductions of plant community mass were greatest at low resource availability. This adds to evidence that herbivore activity poorly predicts herbivore control.