Captions and retention: what works in 2026 (font, case, position, timing)

Captioned short-form video gets watched longer, more often, and to completion at materially higher rates than the same cut without captions. The widely cited industry numbers: 80% of viewers say they're more likely to watch a video to completion if captions are on (Verizon Media / Publicis), and Facebook's own internal tests show captioned ads lift view time roughly 12% on average. That's the reason every retention thread you read this year starts with "burn in your subs."

But "have captions" stopped being the question two years ago. The question now is which captions, and the answers in 2026 are not what they were in 2023.

The lift number, with the asterisk

The 80% completion stat comes from a Verizon Media / Publicis Media study and is the cleanest "captions vs. nothing" headline number around (Newton Tech roundup). 3Play Media's roundup of academic and industry studies adds the supporting context — captioned content gets abandoned less often and watched further (3Play Media). Submagic's own pull across roughly 2 million short-form videos analyzed on its platform is the basis for most of the font and style data below (Submagic).

The asterisk: that lift is averaged across "captions vs. nothing." The marginal lift from going from decent captions to great captions is much smaller, and most of the work below is fighting for that smaller delta.

Font: Montserrat won, even though nobody said it would

In Submagic's 2-million-video pull, Montserrat appears in roughly 1.2 million of them — 61% market share among creators using their tool (Submagic). It is the boring answer and it is the right answer for most niches.

Per-niche, the 2026 picks that actually convert:

• Sports / hype / fitness: Anton. Tall, condensed, hits like a bumper. This is the Hormozi look and it earned its reputation (Submagic).
• Lifestyle / vlog / story: Montserrat or Inter, semi-bold, with a soft outline. Reads as "premium" without trying.
• Finance / educational: Montserrat or IBM Plex Sans. Authority by way of restraint. Avoid anything condensed — finance audiences read more, and tall narrow letterforms slow them down.
• Kids / gaming / MrBeast-coded: Komika (Submagic). Don't use it anywhere else; it'll torch a brand piece in two seconds.

The mistake I see weekly: a finance creator using Anton because their reference deck was Hormozi. The font is doing tonal work the script isn't. Match the font to the niche, not to the creator you wish you were.

ALL CAPS vs. sentence case (the contrarian take)

Here's the take: ALL CAPS captions are losing in 2026, and most creators haven't noticed yet.

The accessibility editorial world has been on this for years — sentence case is faster to read because letterforms have variable height and our eyes use that shape information (NCI, West Coast Editorial). What's new is that creators are catching up. The top OpusClip presets for 2026 ship in sentence case by default, and the highest-retention TikTok styles in Blitzcut's 2026 audit are word-by-word in sentence case with a single highlighted accent word (Blitzcut, OpusClip).

ALL CAPS still wins for sub-3-word punchlines and pure-hype edits. For anything with a sentence in it, sentence case reads ~10–15% faster in eye-tracking work, and on a 30-second short you can feel that.

Use ALL CAPS as a tool, not a default.

Position: lower-middle, not lower-third

The TV "lower-third" convention is wrong for 9:16. TikTok's UI takes the bottom ~20% (progress bar, username, CTA) and the right edge (engagement rail) — captions stuffed into a traditional lower-third get clipped or fight the like button (House of Marketers safe-zone guide).

The position that converts in 2026 is vertical center, slightly low — roughly 55–60% down the frame. That keeps captions in the optical center where the viewer's eye already is (the speaker's face), avoids UI overlap, and is what every Submagic / Captions / Opus default has converged on (Blitzcut).

If your subject is a face, captions go just under the chin. If your subject is a product, captions go where the product isn't.

Word-level highlighting: useful, and overdone

Karaoke-style word-by-word highlighting (one word pops in color, the rest stays white) is the dominant 2026 style, and it works — Submagic, Captions, Opus all default to some flavor of it.

But: highlighting every word is noise. The retention pattern that actually converts is one accent word per phrase, colored or scaled, with the rest in plain white-with-outline. Blitzcut's 2026 audit calls this the "Bold Highlight" — a single key word color-highlighted in yellow, red, or orange — and ranks it the highest-converting style (Blitzcut). When every word fights for attention, none of them get it, and you've also just doubled the visual load on a viewer trying to also watch the subject.

If your tool is bouncing every syllable in three colors, turn it down.

Timing: beats, not words

The "one word per beat" school (each word pops on its own) is built for music-driven edits where the beat is doing the rhythm work. For talking-head content it's a tic — captions out-pace speech and the viewer reads the line before it's said, killing the joke or punch.

The better default for spoken content is natural sentence chunks of 3–6 words, swapped on punctuation and breath, with one highlighted accent word per chunk. Word-level timestamps matter for getting the highlight on the right beat, not for splitting every word into its own card (Vocallab).

Music edits and pure-hype cuts: word-by-word is fine. Talking-head, UGC, ads with a script: chunks.

The recipe most niches should ship

Pulling the data into one default:

• Font: Montserrat SemiBold, 7–8% of frame height
• Case: Sentence case, ALL CAPS reserved for 1–3 word punchlines
• Color: White, 4–6px black stroke, optional 30% drop shadow
• Accent: One highlighted word per phrase in yellow (#FFD60A) or your brand color
• Position: Center horizontally, 55–60% down the frame
• Chunking: 3–6 words per card, swap on natural breath
• Highlight cadence: Pop the accent word on its spoken beat, not every word

That setup is boring on purpose. It's the version that gets out of the script's way.

One takeaway

If you change one thing this week, switch your captions from ALL CAPS to sentence case with a single highlighted accent word per phrase — and ship the same script both ways to A/B it. That's the cheapest retention test in the 2026 caption stack and the one most creators are still leaving on the table.