We present a semantic analysis of the Ramsey test, pointing out its deep underlying flaw: the tension between the “static” nature of AGM revision (which was originally tailored for revision of only purely ontic beliefs, and can be applied to higher-order beliefs only if given a “backwards-looking” interpretation) and the fact that, semantically speaking, any Ramsey conditional must be a modal operator (more precisely, a dynamic-epistemic one).
Thus, a belief about a Ramsey conditional is in fact a higher-order belief, hence the AGM revision postulates are not applicable to it, except in their “backwards-looking” interpretation.