A recommendation engine suggests what a person is likely to want next, learned from patterns in behaviour rather than from a rule somebody wrote. In retail and video it decides which product, which offer, or which branch comes next.
The reason it beats hand-written rules is combinatorial: nobody can maintain "customers who bought X also want Y" across ten thousand products, and the relationships change with season and stock anyway.
The caveat is the cold start, and it is worse in physical spaces than online. Recommendations need history, and a person standing at a screen has none: no account, no session, no past. Systems pitched for in-store use frequently fall back to "most popular", which is a bestseller list wearing an AI label and should be priced accordingly.
Where it genuinely works is where identity already exists: a logged-in customer, a loyalty scan, a personalised video generated from a known account. If a proposal cannot say what it knows about the viewer at the moment of recommending, the answer is nothing, and the recommendation is a guess.