Nobody searches inside their own category
A director briefed on packaging does not look at packaging. They look at sculpture, at hand tools, at how a Japanese bento box solves the same volume problem. Search inside your category and you get the average of your category back.
That is the whole mechanic behind reference work: you are looking for a structural idea in one medium that has never been applied in yours.
The reference shelves that keep showing up
Colour and light: Hockney's flat sunlit palettes, Fincher's cold and deliberate frames. Two opposite ends of the same lever.
Objects and systems: Charles and Ray Eames on how something serious can still be playful, Rem Koolhaas on buildings arguing with their own programme.
Music, more often than people admit. Seventies dance records collided acoustic and synthetic sound in a way that maps almost directly onto mixing photographed and CG elements in one frame.
Where AI fits into the reference stage
Generative tools compressed the gap between a sentence and a picture of that sentence. Prompting a set of style frames to kill three of four directions before anyone books a render node is a genuine cost saving, not a novelty.
The failure mode is obvious once you have seen it: a board built entirely from generated frames looks confident and means nothing. Use it to test a hypothesis, then go find the real reference behind it.
Turn research into something a team can use
A moodboard without annotation is a mood. Write one line under each image saying what you are pointing at — the grade, the lens, the timing, the negative space.
Keep a personal library and revisit it. The best directors I have worked with reuse a small set of obsessions rather than chasing a new one each quarter.