You're severely overestimating the ability of the model to recall a single mostly uninteresting item from it's billions of input documents.