Not the faces, and not the thumbnail on its own. We re-ran our packaging test on 2,779 pairs of real videos from 402 channels picked by a fixed rule, not by us. The videos that beat their channel's baseline did not have more exaggerated expressions, more overlay text or brighter images. What went with them was the package: a clear promise, a title with a pull, and a thumbnail and title that tell one story together. Here is what the replication found, what it ruled out, and how to use it before you publish.
Key takeaways
- On 2,779 pairs of videos from the same channels, the higher packaging score belonged to the better-performing video 55.3% of the time, against 50% by chance. The first, hand-picked study gave 59.5%; expect the lower figure.
- Nine of thirteen factors were confirmed on 1,434 pairs from channels kept aside until the end. All nine read the title, or how the title and the thumbnail fit together.
- No measure of the thumbnail alone was confirmed: not the faces, the gaze, the expression, the overlay text, the brightness or the empty space.
- The gap matters more than the score. A gap of 4 points or more picked the better video in 62.1% of 1,705 pairs; a gap of 1 to 3 points, in 52.4%.
- These factors go with better-performing videos. Nothing here shows that raising one makes a video do better.
What actually goes with a better-performing YouTube video?
On the evidence of this replication, the parts of the packaging that read the title and the match between title and thumbnail. Thumbnail–title coherence came first: in the pairs where two videos differed on it, the better-performing video read higher 59.0% of the time. Close behind were the cliffhanger, curiosity, promise clarity and the withheld answer (what psychologists call the Zeigarnik effect), all between 57% and 59%. Title strength, fear of missing out, rarity and emotion followed, between 54% and 57%.
None of these is large. A factor that goes with the better video 58% of the time leaves 42% of pairs where it pointed the wrong way. That is what packaging evidence honestly looks like: the topic, the timing and the audience decide most of a video's fate, and the package moves the odds a little. Anyone who promises more is selling something we could not measure.
How the replication was built
Our first test, the 321-pair study, used 60 channels we chose by hand. That leaves an obvious doubt: perhaps the result belonged to our choice rather than to the score. So we wrote a second protocol before scoring anything, published it before the factors were chosen and the confirmation half was opened, and took the channels from YouTube's most-popular charts in seven English-speaking regions, by a rule: at least 100,000 subscribers, at least 50 public videos, and none of the first 60. The full method, with the protocol's fingerprint, is on the study page.
Same-channel pairs, as before
Each pair is two uploads from one channel, published close together, where one clearly beat the channel's own recent baseline and the other clearly fell short. Comparing videos inside one channel cancels channel size, audience and design budget, which would otherwise swamp any packaging effect. A score that knows nothing would pick the better video half the time.
Choose on one half, confirm on the other
The 402 channels were split in two before any score existed. On the first half we looked for factors that seemed to go with the better video and nominated thirteen. Then the second half, 1,434 pairs that nobody had looked at, was opened once to confirm or reject each of them, with a strict threshold that allows for testing thirteen things at once. A factor that only fits noise in the first half usually fails in the second, and one did: empty space pointed one way on the first half and the other way on the holdout.
What did not replicate
Three ideas from our first study were tested again, and none held up. A strong exaggerated expression did not go with the weaker video, as the first study hinted: the better video had one in 50.5% of the 719 pairs that differed on it. Overlay text that complements the title did not go with the better video either (51.9% of 804 pairs). A brighter thumbnail came closest, at 52.2% of 2,776 pairs, but missed the corrected threshold.
Faces, gaze and text were not even nominated
The development half looked at every observed aspect of the image: whether there is a face, where it looks, what expression it wears, how much text there is and what it says. None reached the bar to be nominated. That does not prove faces never help. It says that across 402 channels, having one, or a bigger reaction, did not separate the videos that did better from the ones that did not. If you want the full evidence on faces, our article on what the research says about faces in thumbnails collects it.
The thumbnail alone is not where the signal is
An earlier exploratory test pointed the same way: scored with a neutral title in place of the real one, thumbnails ranked their pairs at chance. Our model's own reading of the thumbnail's readability was not confirmed either (52.5%). Readability still matters as a floor, since a thumbnail nobody can read cannot carry anything, but it does not separate a good package from a better one. This matches what we found when swapping titles between videos: the title carries more of the signal than the image.
Why a small score gap tells you nothing
The replication also measured how the size of the gap between two scores changes the answer. When the two scores were the same, the pair counted as a miss. With a gap of 1 to 3 points, the higher score was right 52.4% of the time, barely above a coin toss. From 4 to 9 points it was right 60.0% of the time, and from 10 points on, 65.4%.
| Gap between the two scores | Pairs | Higher score is the better video |
|---|---|---|
| Same score | 162 | counted as a miss |
| 1 to 3 points | 912 | 52.4% |
| 4 to 9 points | 1,063 | 60.0% |
| 10 points or more | 642 | 65.4% |
This is why our analyzer now says "too close to call" when two versions of your packaging are less than 4 points apart, instead of pretending a 2-point rise is an improvement. Two cautions on the table: it was read after the result rather than planned in advance, and the pairs in it performed very differently, while two versions of one video may perform closer together.
How to use this before you publish
- Work on the package, not the picture alone. Write the promise first, then make the thumbnail show the same story the title tells. Coherence was the strongest factor of the nine.
- Make the promise specific. Promise clarity and title strength were both confirmed; a viewer should know what they get before they click.
- Leave one question open. Curiosity, the cliffhanger and the withheld answer all went with the better video, as long as the video then answers the question. Our guide to using the curiosity gap without misleading viewers shows where the line is.
- Do not chase a bigger reaction face. Neither a face nor a stronger expression separated the better videos.
- Compare versions, and respect the gap. Score two versions of your thumbnail and title; under 4 points apart, treat them as tied and let YouTube's own A/B test decide with real viewers. Our guide on A/B testing thumbnails explains how.
Limits of this replication
The channels are established English-language channels that appear in YouTube's charts, which favours large and currently popular channels; nothing here describes a new or small channel. Eleven of the 2,790 pairs harvested were dropped unscored because our analyzer refused their titles, a counting bug fixed since. Views were read once and keep accruing. The study is observational: it shows which readings went with videos that had already done better, not that changing one changes the result. The nine confirmed factors were confirmed once, on one holdout; another sample could rank them differently.
Frequently asked questions
Do faces on YouTube thumbnails get more clicks?
In this replication, faces did not separate the videos that did better from those that did not, and a strong exaggerated expression did not either (50.5% of 719 pairs). That does not mean a face never helps a particular video. It means a face is not, on its own, what better-performing packaging has in common.
What matters more, the thumbnail or the title?
On this evidence, the title and the fit between title and thumbnail. All nine confirmed factors read the title or that fit, and no measure of the thumbnail alone was confirmed. The thumbnail still has to be readable, but that is the floor, not what decides between two good versions.
How big does a difference between two thumbnail scores need to be?
At least 4 points. Below that, the higher score picked the better video about half the time. From 4 points on, it picked the better one in 62.1% of 1,705 pairs, about six times in ten, and from 10 points on, 65.4%.
Is this study independent?
No. We built the score and we ran the test, which is why the protocol was written and published before the confirmation half was opened, the channels were chosen by a rule, and every result is reported, including the ones that went against us. The method and the figures are on the study page.
Can a score predict how many views my video will get?
No. It compares versions of one video's packaging and tells you which reads stronger on the factors that went with better-performing videos. Views depend on the topic, the timing and the audience far more than on the package.

