Literature data were collected on the floristic distribution and toxicity of phytochemicals to herbivores and on herbivore specialization in order to test phytochemical coevolution theory. The theory makes four predictions that can be tested with this information. Herbivores can adapt to novel, more toxic chemicals by becoming specialists, or they can become generalists but at the cost of lower feeding success on any particular host. Thus, the first two predictions are as follows: herbivores should do better on chemicals that are present in their normal host, and this pattern should be stronger for specialists than for generalists. The "escape and radiation" aspect of the theory holds that if a plant taxon with a novel defense chemical diversifies, the chemical will become widespread. Eventually, herbivores will adapt to and disarm it. So the third prediction is that more widespread chemicals are less toxic than more narrowly distributed ones. Because generalists should not do as well as specialists on chemicals disarmed by the latter, the fourth prediction is that the third prediction should be more true for generalists than specialists and should depend on presence/absence of the chemical in the normal host. Multiple regressions of toxicity (herbivore mortality and final weight) on three predictor variables (chemical presence/absence in the normal host, specialism, and chemical floristic distribution) and relevant interactions were used to test these predictions. Chemical presence/absence in the normal host, the interaction between this variable and specialism, and chemical floristic distribution had significant effects on both measures of toxicity, supporting the first three predictions of the model. Support for the fourth prediction (a three-way interaction among all predictor variables) was evident for final weight but not mortality, perhaps because growth is more responsive to toxicity differences than survival. In short, the phytochemistry literature provides broad support for the phytochemical coevolution model.