Languages vary in whether they allow evidential markers in embedded contexts or not. Some languages allow evidentials under attitude predicates, e.g. Cheyenne (Murray 2010) or Cuzco Quechua (Faller 2002). Some languages do not, e.g. Bulgarian, German, or Japanese (Sauerland and Schenner 2007). In the current literature on evidentiality, it is common to attribute non-embeddability of respective markers exclusively to their semantic properties: such evidentials are analysed as speech act modifiers and therefore are expected to be non- embeddable. This expectation rests upon the assumption that speech acts only correspond to root clauses. Krifka (forth.) argues that speech acts can, to some extent, serve as arguments to con- nectives and clause-taking predicates, therefore, are embeddable. I propose that speech acts with evidentials are no exception and formulate semantics for illocutionary evidentials within Krifka (forth.)’s framework. I further argue that (non-)embeddability of evidentials does not stem solely from their illocutionary nature but depends on two factors. (1) The embedding strategy: evidentials across languages are banned from non-finite clauses. (2) The embedder: attitude predicates vary in what can appear in their complements, which sheds light on the behaviour of e.g. epistemic modals (Anand and Hacquard 2013) or shifted indexicals (Sudo 2012). As I will show, it helps to explain the distribution of evidentials as well.