Herbivores may use various cues to detect the presence of anti-herbivore defences in plants. I analyse the effect of such signals on the selection for defences in four game theoretical models. In the models, plants can use either of the two strategies ''defensive'' and ''non-defensive''. The anti-herbivore defence decreases the risk of herbivore attack. Defended plants can produce a signal to inform potential herbivores of their defence status. Herbivores can either recognize the signal without any previous experience or they have to learn to recognize it, and undefended plants can or cannot reduce their risk of herbivore attack by associating with defended plants. The combination of these situations produces four qualitatively different cases. In all four models a signal is a prerequisite for the selection for defence. However, in none of them is it necessary that the signal gives perfect information to herbivores concerning the defence status of the plant Thus, it should not be surprising if herbivores make mistakes and graze well-defended plants. The defensive strategy can only be a pure ESS (evolutionarily stable strategy) when herbivores recognize the signal without previous experience. When herbivores have to learn to recognize the signal the outcome depends on the initial proportions of defended and undefended plants in the population. This means that for a defended mutant to be able to invade a population of undefended plants the proportion of the defended type has to be higher than a certain critical value. Synergistic selection, through non-random association, or convergence of signal type, i.e. Mullerian mimicry, could make selection for defence possible in such cases. In the models where undefended plants can decrease their risk of herbivore attack by associating with defended plants the two strategies may coexist. These are the most likely situations for Batesian mimicry to evolve in, since the ''model'' and the potential ''mimic'' already coexist. Also, in these cases it is not a precondition that the defence and its signal are directly coupled. However, even if undefended plants can reduce their risk of herbivore attack by associating with defended plants, there is a possibility that the grazing cost is increased if undefended plants are found. Such an increase of the grazing cost for the undefended plants diminishes the possibilities for stable coexistence.
