Finding b-roll that doesn't look like stock

Open ten faceless videos about productivity and you will see the same time-lapse of a city, the same hands on a laptop, the same slow push-in on a coffee cup. Not because those are the only clips available, but because everyone searched the same word.

This is a search problem before it is a budget problem. The libraries are much deeper than they look; the queries are much shallower.

Libraries index subjects, not ideas

A stock library does not know what your video is about. It knows what is visible in each clip, because that is what somebody typed when they uploaded it — a woman walking through a train station, a drone shot of a motorway at dusk, close-up of solder on a circuit board.

Your script, meanwhile, is full of concepts. Growth. Pressure. Opportunity. A decision. When you search a concept, the library falls back on whatever clip has been downloaded most by everyone else who searched that concept — which is precisely how twelve clips end up in every video on the internet.

So the translation step is the whole job: for each line of narration, decide what a camera could actually point at.

• “the industry consolidated” → an empty office floor, boxes stacked by a lift
• “demand outstripped supply” → a queue outside a shop before opening
• “he had run out of time” → a departures board flipping over

None of those are literal illustrations of the sentence. They do not need to be. They need to be filmable, findable, and consistent in mood with what is being said.

The b-roll and visual finder does this translation for a passage of narration and returns concrete search phrases — and it refuses to hand back your own abstract nouns as search terms, because that is the failure that makes the whole exercise pointless.

Search the way a shot is described

Three habits find footage that other people are not using.

Search two or three words, not one. “Office” returns everything. “Empty office night” returns a specific kind of shot, from a much smaller pool, which means fewer people are using it.

Include the shot type. Libraries tag aerial, close-up, slow motion, handheld, time-lapse. Adding one narrows the pool enormously and gives you something with a deliberate feel rather than the default mid-shot everyone takes.

Search adjacent, not exact. If the clip you want does not exist, the clip next to it usually does. No footage of the specific machine you are describing? Footage of a similar machine, shot close enough that the difference does not read, almost certainly exists.

Plan the shots before you go looking

The slow way to edit a faceless video is to write the script, then open a library and hunt clip by clip. It is slow because each search is a fresh decision made in isolation, and because you end up choosing whatever is available rather than what the video needed.

The faster order is to break the script into scenes first — where the picture changes and what is on screen for each — and only then search, with a list in hand.

The scene splitter does the break for you and calculates the timings itself rather than asking a model to guess them, so you get a shot list with real durations attached. Then the searching becomes mechanical, which is the point: a list of thirty specific searches is an hour of work, and thirty open-ended decisions is an afternoon.

When the shot does not exist

Sometimes there is no footage, because what you are describing is too specific, too recent, or not visual at all. Three options, in the order they are usually worth trying:

Change the shot, not the script. Ask what else could be on screen during that sentence. A line about a company’s collapse does not need footage of that company; it needs something that carries the feeling of the sentence.

Use something other than footage. A still with a slow push, a simple animated chart, text on a plain background. Faceless videos over-rely on moving footage, and a plain card can be a relief in a sequence of clips.

Generate it. AI video has become genuinely useful for short abstract inserts — a texture, a shape, an impossible camera move. It is still poor at people, hands and any legible text, and a three-second insert is a far better use of it than a hero shot. Costs vary sharply by model and length; the AI video cost calculator prices a specific amount of generated footage rather than a monthly plan, which is the number that actually matters when you are deciding between a clip and a subscription.

Keep track of where clips came from

This is dull and it matters. If a channel is monetized, every clip in it needs a licence that permits monetized use, and “I found it on a free site” is not a record.

Keep a text file per video: clip, source, licence, date. It takes seconds per clip and it is the difference between answering a claim in a minute and re-editing a video you published a year ago. Licences also change — a library can alter its terms, and what you can prove is what you downloaded under.

The short version

• Translate concepts into things a camera can point at, before you search.
• Search two or three concrete words plus a shot type.
• Break the script into scenes first, then search from a list.
• If a shot does not exist, change the shot rather than the sentence.
• Record the licence for every clip while you are downloading it, not later.