Linguistic manipulation works the normal way, which is by representing the world a given way, and people then responding rationally by behaving based on the representation.
This is why priming studies don't replicate, and it's why the strongest subliminal messaging effect you'll see on stage is getting people to say random words and numbers, i.e., bending trivial and arbitrary behavior in situations where the "manipulated" agent isn't deeply invested. The video we're discussing, which droid chocked up to "NLP" and subliminal messaging, is a confirmed stage trick and never happened.