An RCT tests only the latter, but all too often ambiguous language is used to imply that the results then say something general about the first two (and - perhaps not so coincidentally - extend the range of authority claimed by the speaker).

This is the crucial limitation of RCTs for this category of problems - they are big, they are expensive, they answer only a small number of questions (very often only one) and if you don't see an effect, often can’t tell you why - a significant drawback vs mechanistic approaches.