Which caption styles work on talking-head videos?

The short version

Most big differences weren't the captions

Creators who use a style also tend to make a certain kind of video, for a certain kind of audience. So we checked every result three ways, each fairer than the last, using 11,525 videos from 1,026 creators labelled for what each video is.

1. All creators. The first, raw comparison between videos with the style and without it.
2. Same type of video. Teaching against teaching, story against story, with the hook, setting and production held equal too.
3. Same creator. Each creator's videos with the style against their own videos without it. Their audience, niche and skill cancel out.

Then one more check: does the "effect" also show up on the creator's previous post? A caption can't change a video made before it, so if it does, the cause is timing or circumstance.

"Holds up" = clear within the same creator after correcting for testing many styles at once, with the previous-post check flat. "Probably real" = clear within the same creator at the usual 95% level but not after that correction. Bars run from −50% to +50%.

Every caption choice, same creator: with it vs without it

Every dot is how a creator's videos with that choice did against their own videos without it. The shaded band is ±10%. Dots near zero with short lines mean the choice made no real difference.

Saves and shares are TikTok only (Instagram doesn't report them). Views are per video; likes, comments, shares and saves are per view. Hover a dot for the exact numbers and how many creators used both options.

Words from the first frame, out of the app's way

Measured from on-device text recognition of the first 8 seconds, six frames a second. App zones: the bottom 20% and the right-hand button column on TikTok and Instagram, and the top 9% for the tabs.

Breakout odds for each caption style

Each dot is how much more (right) or less (left) likely a video with that style was to break out, compared with videos without it, after accounting for niche, platform, follower count, post age and video length. The line is the 95% confidence interval. Lines that cross the dashed 1.0 line mean no clear difference.

Every caption style, with real examples

Each style shows real frames from videos in the study. "Used by" is the share of creators using the style in a random sample of their recent videos; "of breakouts" is the share among videos that hit 3x the creator's usual views.

Hook headline, no captions

No clear effect

A short headline sits on screen (usually at the top). The speech itself is not captioned.

23%used by (all talking-head videos)

25%of breakout videos

+0.7 ptsvs 6–12 months earlier

1.11 (0.99–1.25)breakout odds (95% CI)

Full-line captions

No clear effect

A whole line or sentence of speech at a time, all words the same colour.

15%used by (all talking-head videos)

15%of breakout videos

-2.0 ptsvs 6–12 months earlier

1.01 (0.88–1.16)breakout odds (95% CI)

2–3 words, no highlight

No clear effect

Short chunks of two or three words, all the same colour.

10%used by (all talking-head videos)

9%of breakout videos

-1.4 ptsvs 6–12 months earlier

0.95 (0.80–1.14)breakout odds (95% CI)

One word at a time

No clear effect

Each spoken word flashes up on its own, in normal case.

7%used by (all talking-head videos)

8%of breakout videos

+2.0 ptsvs 6–12 months earlier

1.20 (1.00–1.45)breakout odds (95% CI)

One word at a time, ALL CAPS

No clear effect

Each spoken word on its own, in capitals.

5%used by (all talking-head videos)

5%of breakout videos

+1.5 ptsvs 6–12 months earlier

1.04 (0.86–1.25)breakout odds (95% CI)

Karaoke colour highlight, ALL CAPS

No clear effect

Two or three words in capitals; the word being spoken turns a colour (yellow, green, pink).

7%used by (all talking-head videos)

7%of breakout videos

-0.9 ptsvs 6–12 months earlier

0.94 (0.77–1.16)breakout odds (95% CI)

Karaoke box highlight, ALL CAPS

No clear effect

Two or three words in capitals; a coloured box sits behind the word being spoken.

4%used by (all talking-head videos)

5%of breakout videos

-1.0 ptsvs 6–12 months earlier

1.20 (0.96–1.49)breakout odds (95% CI)

Typed-on captions

Leans fewer breakouts

Words are added one by one as they are spoken, so the line builds up like typing.

10%used by (videos with captions)

7%of breakout videos

-7.5 ptsvs 6–12 months earlier

0.73 (0.57–0.93)breakout odds (95% CI)

Only key words coloured

Leans fewer breakouts

Most words plain; a few chosen words get a colour or size, not every word in turn.

10%used by (videos with captions)

6%of breakout videos

+2.9 ptsvs 6–12 months earlier

0.67 (0.48–0.94)breakout odds (95% CI)

Tall bold font with pink highlight

Leans more breakouts

A tall, narrow font (like Anton or Bebas) with a pink box or colour on the spoken word: the CapCut pink-box look.

4%used by (videos with captions)

5%of breakout videos

-0.8 ptsvs 6–12 months earlier

1.36 (1.03–1.80)breakout odds (95% CI)

Blue highlight

Leans fewer breakouts

The spoken or key word is marked in blue.

4%used by (videos with captions)

4%of breakout videos

+1.5 ptsvs 6–12 months earlier

0.73 (0.56–0.97)breakout odds (95% CI)

Serif captions

No clear effect

White captions in a serif typeface, often editorial or italic.

5%used by (videos with captions)

5%of breakout videos

+1.9 ptsvs 6–12 months earlier

1.08 (0.81–1.44)breakout odds (95% CI)

Examples are frames from public posts, shown for research and commentary and linked to the original. Creators can ask us to remove theirs at yapanddone.com/contact.

What's rising and what's fading

Change in the share of creators using each look, compared with the same niches 6–12 months earlier.

How creators caption, choice by choice

Share of creators making each choice, split by platform, with the breakout odds for that choice. Caption choices are measured among creators who use captions.

Niches use different looks, and fitting in matters

Which caption setups are most common differs a lot between niches. No style did better in one niche than in another, but using a setup that's out of place in your niche cost views.

Small accounts feel it most

The main results again, within each follower band. Words on screen helped saves at every size; going without text, and looking out of place in the niche, cost small accounts the most.

Green or red = the 95% range excludes zero. "–" = fewer than 30 creators in that band used both options.

Most captions look like CapCut, and the tool doesn't matter

Same captions, different headlines

Instagram creators more often pair captions with a headline (62% vs 52% on TikTok) and less often use a headline alone (13% vs 25%). Caption details (words on screen, highlight, font, case) did not differ significantly between the platforms. The Instagram sample is smaller (161 creators, about 120 breakouts), so its numbers are less certain.

Claims we'll re-test next quarter

On November 7, 2026 we wrote these results down before collecting the next data. Next quarter we test each one on new videos only and publish every outcome, including the ones that don't hold up. Each claim is re-tested on posts collected at the next quarterly refresh, leaving out every post used here, with the same code. Same creator, same controls (type of video, hook, production, setting, post age, length). A claim replicates when the new estimate points the expected way with p < 0.05. Every claim is reported, whether it replicates or not.

All registered claims, for both studies: preregistered.json (SHA-256 b70ff7763879a253…).

How talking-head creators caption, tracked every quarter

The share of creators doing each thing this quarter, all ten niches weighted equally, and the change compared with posts from the same niches 6–12 months earlier. Next edition: Q1 2027. Every edition is in index.csv.

Green or red = moved 2 points or more. "New" = first measured this edition.

What to do with this

How we did it

• Creators: in each niche we found English-language creators through keyword and creator search, then had a language model screen every account for being genuinely in the niche (score 60 of 100 or more) with at least 2,500 followers. Instagram creators were found through search and by checking the same handles TikTok creators use.
• Videos: we collected each creator's latest posts and kept videos 8–180 seconds long, posted 7–60 days before collection (7 days so views had settled; 90 days on Instagram, where creators post less). Paid or boosted posts were excluded.
• Breakouts: a breakout is a post with at least 3x the views of the creator's typical post, measured against the median of their roughly 20 nearest posts in time. The comparison group is a random sample of the same creators' posts, not filtered on performance. At most three posts per creator in each group.
• Talking head: Apple Vision face detection kept videos where a face fills the frame in at least half the sampled frames; the labelling model then confirmed a person speaking to camera carries the video (including split-screen and picture-in-picture).
• Labelling: from each video we took the first 8 seconds at 6 frames per second plus one frame at 60% of its length, picked the 2-second stretch with the most caption activity, and showed Gemini 3.8 Flash three images: four moments across the video, a 12-frame burst stamped with times, and one full-resolution frame. It answered a fixed codebook of 24 questions.
• Quality checks: 48 videos were labelled independently by separate raters (agreement shown below); 211 were labelled twice to check consistency; every label was cross-checked against on-device text recognition. Fields with less than 85% agreement are reported in coarser groups.
• Cause or coincidence: 11,525 random posts from 1,026 creators (about 11 each) were also labelled for what the video is: its type (teaching, list, story, opinion…), hook, production, setting, b-roll and framing. Each result was checked three ways: between creators with all of that held fixed; within each creator, comparing their posts with a style against their own posts without it; and on the creator's previous post, which a caption cannot affect (a pattern there means timing or circumstance, not the caption). We also report how strong a hidden factor would need to be to explain a result away (the E-value).
• Timing, placement and pace: from the on-device text recognition of the first 8 seconds (6 frames a second). Text that stays on screen 1.5 seconds either side is the headline; the rest is captions. App UI zones: the bottom 20% and the right-hand button column (x > 85%, 40–80% down) on TikTok and Instagram, and the top 9% for the tabs. Talking pace counts each caption word once, when it first appears, and is compared within each caption size.
• Consistency: each creator's share of random posts using their most common caption setup, compared between creators with niche, platform, follower count and posts sampled held fixed. Follower counts at collection are stored so follower growth can be measured at the next edition.
• Statistics: "used by" is the share of creators using a style in the random sample, each creator counted once and each niche weighted equally. Breakout odds come from a logistic regression of breakout vs comparison posts with controls for niche, platform, follower count, post age and video length, and standard errors clustered by creator. With about a hundred styles tested, we apply a Benjamini–Hochberg false discovery correction (10%): "proven" requires passing it; "suggestive" clears the 95% interval but not the correction. Trends compare the same niches against posts from 6–12 months earlier.

How reliable the labels are

Limitations

• This is observational data, not an experiment. Comparing creators with themselves and holding the type of video fixed removes most of the obvious explanations, but a factor we did not measure could still play a part.
• Only text burned into the video is measured. TikTok and Instagram can also show their own captions, which do not appear in the video file.
• Labels come from still frames, so fast animation details can be missed; font identification is the least reliable field (77% agreement in three groups).
• Talking pace is read from the captions, so it covers captioned videos only and is noisy; consistency is compared between creators, not within them, so it is a correlation.
• Instagram coverage is thinner than TikTok, and creators were found through search, so very small accounts are under-represented.
• Results describe English-language talking-head videos in these 10 niches in late 2026; styles change, and we plan to refresh the study.

Questions people ask

Use it, cite it

The aggregated results are free to reuse under CC BY 4.0. Journalists, researchers and AI agents can pull the full numbers below; the plain-text summary contains the whole study.

Cite as

Copy citation

Press questions or a custom cut of the data (a niche, a platform, a time window): contact us. We refresh this study as styles change; this page always shows the latest snapshot.