Describe what the camera sees, not what the shot is for. That means the framing, how many people are in it, what they're doing, what's behind them, and any text on screen. Then run that description against footage that has been indexed for its picture. Filenames and transcripts won't carry it. Transcript search does little for most b-roll because b-roll rarely has useful speech. Folder names and tags only help if someone logged the shot, in the words you'd use now, when it came in.
Why b-roll is the hard case
Talking-head footage can be searched by what people say. B-roll mostly can't. Product close-ups, hands at work, a model walking past a window: these clips usually have ambient sound or nothing, carry camera-generated filenames, and sit in folders named after the shoot date. What makes the shot useful is only visible in the picture.
The routes
Scrubbing and memory
Someone remembers roughly which shoot the shot came from and scrubs through it. This works for recent shoots and small libraries. The cost is time plus reliance on one person's memory, and it stops working once the library outgrows what anyone remembers.
Logging and tags at ingest
Clips get keywords or bin names when they arrive ("kitchen, morning light, product in hand"). Search then works as long as the words you search with match the words someone logged. The costs are someone logging every clip, tags that drift between people, and questions nobody thought of at ingest. Tools that apply tags automatically take over the logging, with tradeoffs of their own, covered in automatic tagging.
Transcript search
This is strong for interviews and voice-over, where the words are the content, as in searching by spoken words. For silent b-roll it finds nothing, because there's nothing to transcribe.
A visual search layer
Footage is uploaded to a project and indexed once for picture and sound. You type a description and get back time ranges to review. The costs are the upload, a search allowance, and reviewing every result, since results are candidates. In Vivu, a description like "hands opening a box on a light wood table" comes back as openable time segments, each with a short reason describing what's in frame, so you can often tell from the reason whether it's the shot before you open it.
How to write the description
Name what's visible. Posture, action and framing do most of the work. "Standing, holding a microphone, gesturing" will find the same person each time they're shot that way, even across a long recording. "The host" won't, because nothing in the picture says who the host is.
Describe people by appearance, not identity. A search layer can describe what someone looks like and what they're doing, but it won't know who they are. Searching for your founder by name fails where "a person with short gray hair and a navy jacket at a desk" can work.
Add composition when it matters. How many people, where they are, and what's behind them all narrow the result, and a good result's reason will line up with each detail you wrote. Start with the main features, though. A long list of details can rule out shots that are close enough to use.
Treat an empty result as two possible answers. The shot may not exist: an event filmed mostly wide may have no single-person close-ups. Or the description may be too strict. Loosen one detail at a time and search again, so you can tell which it was.
Leave handles. Returned segments vary in length, and some are only a few seconds of the action. Plan on extending them a little on each side in the edit.
Search before you book the shoot
The most useful time to run a shot search is before a reshoot. If the next brief calls for a product-in-hand shot or a slow push on packaging, search the existing footage first. Sometimes the shot is already there from an earlier campaign, in a take nobody used.
When you don't need this
If the library is small, one editor logged it carefully, and that editor is still on the team, tags will answer most questions. If your b-roll comes from stock libraries, their own search already works by description. If you shoot fresh for every campaign and never reuse footage, there's nothing to search.
Which side you're on
The test is simple. When you need a shot, can you name the file, or can you only describe it? If you can usually name it, or your tags match how you ask, keep that system. If you often know exactly what the shot looks like and still can't find it, the gap is between your description and how the footage was indexed, and closing it takes search that reads the picture.
FAQ
Can I find b-roll of a specific person by describing them?
You can describe what the person looks like and what they're doing, and a visual search can match that. It won't identify who the person is, so searching by name doesn't work. Write the visible details: clothing, hair, setting, action.
Why does my shot search return nothing?
Either the shot isn't in the footage, or the description is stricter than any real shot. Remove one detail at a time and search again. If broader versions return nearby shots but never the one you pictured, the shot probably was never filmed.
Does transcript search work for finding b-roll?
Only when the b-roll has speech or voice-over that describes it. Most b-roll is silent or ambient, so transcripts miss it. Text that appears on screen is part of the picture and needs a search that reads the picture.
Is it better to search for "wide shot" or to describe the whole scene?
Describe the scene when you know it. Framing alone ("wide shot") matches every wide shot, while adding the people, setting and background narrows the results to the one you mean. If a detailed description returns nothing, move back toward the shorter version.