Good content research is mostly ruling things out. Here is one popular theory we checked and had to drop
We checked whether caption length moves reach across 2,308 Reels. It does not. What research should actually look at instead. 25 September 2026.
What the numbers say
The useful half of research is finding out what does not matter, because that is what stops you spending next month on it. Caption length is the clearest example we have. Across 2,308 Reels sorted by how long the caption runs, reach barely moves. Captions under 60 characters reached a median of 0.56 times their follower count. Captions of 350 characters or more reached 0.48. In between, 60 to 149 characters reached 0.50 and 150 to 349 reached 0.52, which is not even in order. The whole range from shortest to longest is under a tenth of a view per follower, and it does not trend. For comparison, on the same corpus the measured quality of a Reel's opening separates 0.35 from 1.37. Caption length is noise and the hook is not, and no amount of writing longer captions will close that gap.
Research a claim before you build a habit on it. Caption length is easy to measure, easy to change and widely advised, and in this corpus it did nothing. Those three things together are what a testable myth looks like.
Sample of 2,308, measured on 2026-09-25. Scrptly production database, queried 2026-09-25. 2,308 Reels deduplicated by shortcode, each with a caption, a view count and its account's follower count. 11 Reels carrying no caption at all were excluded as too few to report.
The measurement
Caption length in characters, including spaces and hashtags. Views per follower is the Reel's views divided by its own account's follower count. The four bands hold all 2,308: 340 plus 516 plus 556 plus 896. A further 11 Reels carried no caption at all and are left out, because eleven is not a sample. Read the median column and notice it does not go anywhere.
| Caption length in characters | Reels measured | Median views per follower | Top tenth reaches |
|---|---|---|---|
| Under 60 | 340 | 0.56 | 6.64 |
| 60 to 149 | 516 | 0.50 | 10.97 |
| 150 to 349 | 556 | 0.52 | 9.14 |
| 350 or more | 896 | 0.48 | 9.29 |
One Reel, beat by beat
What research should look at instead. One real Reel from the corpus, not a composite: 399,441 followers, seen 11,947,227 times, which is 29.91 times the following. It runs 26 seconds and its caption is 506 characters, which sits in the band that did slightly worse than average. The caption is not why it travelled. The structure below is, and it is unusual: there is no argument in it at all, only evidence, stacked.
| At | What happens |
|---|---|
| 0:00 to 0:02 | Text on screen plus the single most surprising clip. The proof arrives before the premise |
| 0:02 to 0:10 | Four more clips in quick succession. Different settings, different people, the same reaction |
| 0:10 to 0:18 | Three more, now shot close: hands, faces, laughter. The evidence gets more intimate rather than more varied |
| 0:18 to 0:22 | Pattern interrupt on an unexpected angle, which resets attention before the end |
| 0:22 to 0:26 | Close on the ordinary version of the same thing, so it lands as something that happens to everyone |
What to do next
- Write down what you currently believe moves your reach, then check whether any of it has ever been measured on more than one account. Most posting advice has not.
- Start with the claims that are cheap to test and expensive to follow. Caption length is the archetype: it costs real effort every post and, in this corpus, changed nothing.
- Do the comparison inside one length band or one account wherever you can. Most apparent findings in short form are the length of the video wearing a disguise, and controlling for it makes about half of them disappear.
- Research other people's Reels rather than only your own. Your own account gives you one creator's habits; a corpus gives you the spread, and the spread is where a claim either survives or does not.
- One honest limit that applies to everything above: these are Reels from accounts somebody chose to analyse, so the corpus skews towards content that already worked. It is good for telling strong Reels apart from each other and poor for telling you how often something fails outright.
- Scrptly is the research tool built for this: paste a handle and it returns the transcript, the hook score and the beat order for the Reels that travelled. The first analysis does not need an account.