Two to four words a card, and check it on a bus

A comfortable spoken read runs at about two and a half words per second. Silent reading of short text is faster. That gap is why burned-in captions work at all.

The margin disappears when too many words appear at once. A full sentence of fourteen words demands a sustained read. While the viewer is reading they are not watching your face or your product.

Two to four words at a time tracks speech naturally. It leaves the eye free. Longer groupings work in slower editorial pieces. In a fast vertical ad they cost you the picture.

There is a second effect people miss. Words appearing in rhythm create motion in an otherwise static shot. That holds a scrolling eye on its own.

A caption that changes every beat does part of the job you were about to hire b-roll for.

Captions get read on a small screen, in daylight, on a moving vehicle, over a video that is itself moving. That calls for heavy weight, high contrast, and something separating the type from what is behind it.

Mid-weight type over a bright, busy background is the most common legibility failure. It happens because it looked fine on a large monitor in a dark room.

Test against the worst frame, not the average one. One second of white sky behind white text is enough to lose the sentence. That second is usually your hook.

Position matters as much as the type. Platform interface elements sit across the top and bottom. A caption pinned to the lower edge ends up under a username.

Slightly above centre is a safe default. It survives next year's app redesign. Then play the export on a phone. Arm's length, half brightness, outdoors.

Highlighting a word makes it pop. Highlighting every third word makes nothing pop and turns the ad into a ransom note.

Treat emphasis as a small allowance per ad. Spend it on the words that carry the argument. The number, the objection, the surprising claim. Emphasise verbs and adjectives and you are adding motion, not meaning.

The same restraint applies to animation. Word-by-word appearance is fine. Words that bounce, spin and change colour compete with your message. Your message loses.

Style carries a signal of its own. A heavy karaoke pack says this is a fast entertainment clip. A quiet lower-third says this is a corporate explainer.

A mismatch between caption style and content reads as trying too hard. It lands in the first half second.

Pick one pack per campaign and keep it. Switching every ad costs you the small recognition benefit of looking like yourself. It also means you never learn whether style mattered.

One mis-transcribed word can turn a claim into a nonsense sentence. Product names are exactly what automatic transcription gets wrong.

It will confidently render your brand as a similar-sounding common word, in the frame where you needed the name to land.

So read the captions before you export. Every time. It takes thirty seconds and it catches the error that would otherwise run for a week.

Build a short list of the words your transcriber gets wrong. Brand names, ingredient names, anything with an unusual spelling. Check those first and you catch most of it in ten seconds.

This is not a criticism of any one tool. It is the state of the technology. Any page telling you otherwise is selling something.

In Cutroom the captions come from your own audio, timed to the word without anybody dragging anything.

Correct a mis-heard product name in the transcript. The correction rebuilds the cut. That is the difference from a caption app, where fixing a word means re-timing every word after it.

Switch the style, colour, weight, size and position across six shipped packs. clean-ugc, bold-impact, karaoke-flow, editorial, fast-native and minimal-lux. Every one is built for a small screen in daylight.

Emphasise a word and it pops in the captions. That is how you spend the two or three word allowance from the section above, without hand-styling anything.

The facts to plan around. Six packs is what ships, so a caption set in your own licensed typeface comes from elsewhere. There is no timeline for per-word positioning. The output is a 9:16 MP4.

Every transcriber mishears a product name on the first pass. What differs here is what the fix costs. One keystroke in the transcript, and the cut rebuilds around it.