We use GPs for our own Bayesian optimization of search engine parameters for relevance. I suspect it could also be a backbone of doing active learning to learn about the search or recos click data’s blind spots. (Basically also Bayesian optimization, but a stronger explore bias)