Why do some Reels hit a million views and others not?
According to Kaufman Lab data as of September 2026, of the 33 editing and script features, only one is linked to views, and only for the second creator: a permutation test on 265 videos by two creators found nothing for the first creator and a repeated format for the second. Topic is what moves views: the median for family videos is 712,777 against 291,673 for videos about the creator's craft, a 2.4× gap. Techniques affect comments instead.
- Sample
- 265 videos · 2 accounts · published 2023–2026 · 33 features of speech, shot, caption, sound and feed · permutation test with 20,000 shuffles, threshold p < 0.0015 · 49 second-by-second breakdowns read closely, 423 images checked by a second agent
- So what
- Copying editing techniques is fine as basic hygiene, but it does not explain reach. Over three years the creator more than doubled the number of cuts per video, yet median views fell: 477,000 in 2023–24, 419,000 in 2025, 343,000 in 2026. What separates a 5-million-view video from a 150,000-view one is who it speaks to: in the top videos the viewer feels recognized, not taught.
Published Sep 2026 · data as of Sep 2026 · assembled by 5 workflows and 45 AI agents, 84 cross-checks · statistics recomputed by script · reviewed by a human
How to pick a video that will take off: six steps
- Topic and audience first. Viewers should recognize themselves, not feel sent back to school. This decision is made before shooting.
- Check the promise. A video sells one of three things: a name for how the viewer feels, permission not to do something, or a mistake someone else has already paid for.
- Techniques are hygiene, not the engine. A three-second hook, speech pace, pauses and length did not explain views: 265 videos, no link.
- Scene before the general point. An abstraction has to turn into evidence in the shot, or the viewer won't retain it.
- Polish for 30 minutes max. More does not pay off while the topic is chosen at random. Put the time saved into the next video.
- Track comments separately. A CTA at the end drives discussion, not views. These are different goals and different videos.
Which approach for which goal
- You need reach — change topics, not editing: the topic is visible in the first seconds, the editing is not.
- You need discussion — a CTA and a question to the viewer at the end.
- You need sales — first name a feeling the viewer recognizes, and only then the product.
Below is the full analysis: 265 videos by two creators, 33 features and a permutation test.
What we found: topic drives views, techniques drive comments
Techniques do not explain views. A three-second hook, showing what you talk about, speech pace, pauses, video length, a loop where the hook word returns at the end: none of the 33 features passed the significance threshold on the first creator's 106 videos. For the second creator one feature passed: the number of videos made in the same format, r = +0.26. The main confirmed driver of views is topic.
Comments can be explained, but differently for each creator. For the first creator, videos with a call to action (CTA) at the end get 16.6 comments per thousand views against 0.9 for the rest. In ten cases out of twelve the CTA asks viewers to DM a word, and people post that word under the video instead. For the second creator four of five features have a negative effect: the faster the speech, the more twists, the choppier the editing and the more literally the shot repeats the speech, the fewer comments a video gets.
The mechanism shows only when techniques combine. A close reading of 49 second-by-second breakdowns found what a feature-by-feature analysis could not: all 49 top videos have a plot twist, a concrete scene comes before the general point in 88% of them against 38% of control videos, and every abstraction in speech turns into visible evidence in the shot.
What we tested: seven of the creator's rules, turned into numbers
In two videos the creator explains how to make videos that take off. We turned each rule into a measurable feature that can be extracted from a second-by-second breakdown: "the hook decides in three seconds" became eight features of the opening seconds, and "show everything you talk about" became the share of seconds in which the named object is in the shot. The significance threshold was set before counting: 0.05 divided by 33 tests, or p < 0.0015.
How we counted: breakdowns, features, permutation test
Each video was broken down second by second: speech, shot, caption, volume. We extracted 33 features from these breakdowns, measured their link to views and comments across the whole corpus with rank correlation, and tested significance with a 20,000-shuffle permutation test, dividing the threshold by the number of tests. Separately, the 49 breakdowns of the top videos were read as texts: 606 moves were reduced to 20 types and 5 structures, and each of the 423 images was checked by a second agent.
| Lab practice | What it gave this study | Canonical name · source |
|---|---|---|
| Significance by shuffling, not from a table | The corpus is small and the view distribution is heavy-tailed; 20,000 shuffles give a p-value with no assumption of normality | Permutation testFisher, "The Design of Experiments", 1935 |
| The threshold is divided by the number of tests | 33 features: without correction one or two would pass by chance; with it, none passed for the first creator and one for the second | Bonferroni correctionDunn, Journal of the American Statistical Association, 1961 |
| Ranks instead of values | Views range from 5 thousand to 5 million; ranks keep a single video from outweighing the corpus | Rank correlationSpearman, American Journal of Psychology, 1904 |
| Correlations across the whole corpus, not "top vs bottom" | Comparing the extremes once showed a tenfold difference that did not exist | Selection on the dependent variableGeddes, Political Analysis, 1990 |
| Every image checked by a second agent | 423 images from 49 breakdowns; disputed ones removed | Independent double codingKrippendorff, "Content Analysis", 2004 |
Results: two drivers and one mechanism
| Author's rule | Measured feature | Views | Comments |
|---|---|---|---|
| Hook wins in 3 seconds | 8 first-seconds features | not sig. | direct "you", 2nd creator: r = +0.26 |
| Show what you talk about | speech–shot match | near miss: r = +0.20 | 2nd creator: r = −0.29 |
| Repeatable format | videos in same format | 2nd creator: r = +0.26 | not sig. |
| Emotional delivery | pace, pauses, volume | not sig. | 2nd creator: words/s r = −0.31 |
| Longer videos win | time | not sig.: r = +0.03 | 1st creator: r = +0.30 |
| CTA at the end | yes or no | not sig. | 1st creator: r = +0.49 |
| Video topic | family or craft | ×2.4, p < 0.0001 | not tested |
How far to trust it: two creators, not the market
- Sample
- 265 videos by two creators in one niche. Conclusions about features apply to these accounts; a third account needs its own measurement, since the same feature works in opposite directions for the two creators.
- Spread
- Views range from a few thousand to 5 million, so medians and ranks are used throughout, not means. Recent videos are still gaining views, so the year-by-year comparison used only videos older than one month.
- Does not generalize
- The first creator's comments depend on a CTA that asks viewers to "DM me a word": the mechanism is incidental and may not work in other niches.
- Not measured
- Retention and completion rate: the accounts belong to other people, so only public views, likes, comments and duration are available. We do not publish the corpus's total reach until it is checked against the source data.
- What has already changed
- The first version of the analysis compared top videos with bottom ones and found a tenfold difference in editing pace; recalculation on the whole corpus did not confirm it. This version measures links only across the whole corpus.
What to do: topic first, techniques second
The first decision about a video is its topic and audience: a viewer who feels recognized, not lectured. Treat editing techniques as hygiene: a scene before the general point, a plot twist, an abstraction turned into visible evidence. Don't spend more than 30 minutes polishing a video until you have six videos in one format to measure.
A video does not sell information. It sells one of three things: a name for how the viewer feels, permission not to do something, or an option the creator has already ruled out by failing at it.
Common questions about Reels
What is feature-based Reel analysis?
Feature-based Reel analysis is a Kaufman Lab method: each video is broken down second by second into speech, shot, caption and sound, measurable features are extracted from that breakdown, and their link to views and comments is tested statistically. On 265 videos by two creators, the method showed that topic drives views, not editing techniques. It does not measure retention: only the account owner can see that.
Why don't my Reels get views when the editing is good?
Because editing is not linked to views. On the first creator's 106 videos, none of the 33 features of hook, pace, captions and length passed the significance threshold. Over three years the creator more than doubled the number of cuts per video, while median views fell from 477,000 in 2023–24 to 343,000 in 2026. Only topic turned out significant: family videos get 2.4 times more views than craft videos.
Is it true that the hook in the first three seconds decides everything?
Not according to data from these two accounts. Eight features of the opening seconds (caption, question, a number with stakes, direct "you" address, brand name) show no link to views. One is linked to comments: direct "you" address in the second creator's hook, r = +0.26. You still need a hook, but it is not what separates a million views from a hundred thousand.
What Reel length is best for views?
Length is not a lever: the correlation with views is r = +0.03 at p = 0.76. There is a link with comments: the first creator's longer videos with dense captions get discussed more, r = +0.30. That creator's videos got about 1.6 times longer over three years with no growth in views.
How do you get more comments on a Reel?
For the first creator, videos with a CTA at the end get 16.6 comments per thousand views against 0.9 without one. In ten cases out of twelve the request is "DM me a word", and people post the word under the video. For the second creator, comments grow when the video is not overloaded: fewer words per second and twists per minute, steadier editing, and a shot that does not literally repeat what is said.
What sets a creator's million-view Reels apart from the rest?
A close reading of the 49 breakdowns of top videos showed three things. All 49 have a plot twist. A concrete scene comes before the general point in 88% of them against 38% of control videos. Every abstraction in speech becomes visible evidence in the shot: "wrung out" is shown as a woman sitting on the floor with her back to the door. The useful advice in top videos is ordinary; what sells is a name for how the viewer feels, a permission, or an option the creator has already ruled out by failing at it.
How can I check whether a technique works on my own account?
Take at least 6 videos in one format and compare their median views with the account median. Measure the link across the whole corpus with rank correlation, test significance with a permutation test, and divide the 0.05 threshold by the number of features tested. Do not compare only top videos with bottom ones: that once produced a tenfold difference that did not exist.
- Public pages of two Instagram accounts and their creators' open videos about how they shoot; data on views, comments and duration were collected from the public API on 28.08–04.09.2026. The account addresses are not published: the analysis is about mechanics, not people.
- Fisher R. A. The Design of Experiments. Oliver & Boyd, 1935 (permutation test).
- Dunn O. J. Multiple Comparisons Among Means. Journal of the American Statistical Association, 56(293), 1961 (Bonferroni correction).
- Spearman C. The proof and measurement of association between two things. American Journal of Psychology, 15, 1904.
- Geddes B. How the Cases You Choose Affect the Answers You Get: Selection Bias in Comparative Politics. Political Analysis, 2, 1990.
- Krippendorff K. Content Analysis: An Introduction to Its Methodology. Sage, 2004.
- Kahneman D., Tversky A. Judgment under Uncertainty: Heuristics and Biases. Science, 185, 1974 (on why comparing extremes misleads).
- Fluxnote. Instagram Reels benchmarks: how many views a follower base delivers. fluxnote.io
- Socialinsider via Superdirector. Follow conversion rate for Reels and Shorts, 2026. superdirector.app