Open Measurement Study · October 2026

The Thumbnail Truth

An entire industry sells thumbnail optimisation. We measured 440 of them across two and a half years to find out whether the things those tools measure actually predict anything. They mostly don't.

Panos Khan Foundation 440 thumbnails · 11 quarters Data & method published in full

<2.5%

of the variation in how a video performs is explained by any pixel-level property of its thumbnail we could measure — and the one effect that survived correction turned out to be channel size in disguise.

What we did

Method

We took one niche — “AI tools” — and sampled the 40 most-viewed videos per quarter for eleven consecutive quarters, from the start of 2024 to the third quarter of 2026. Only standard-length videos (4–20 minutes) were included, so every thumbnail is the same 16:9 format and Shorts don't contaminate the sample. That gives 440 thumbnails, evenly spread across time.

Each thumbnail was downloaded at full resolution and measured on seven properties, computed from the raw pixels:

Brightness — mean luminance
Contrast — standard deviation of luminance
Saturation — mean HSV saturation
Colourfulness — the Hasler–Süsstrunk metric
Edge density — mean Sobel gradient magnitude, a proxy for visual busyness and text load
Skin-tone fraction — share of pixels in the YCbCr skin range, a proxy for a face or person being present
Dark fraction — share of pixels below a luminance threshold

Then we asked two questions. First: do these properties drift over time? Second, and more usefully: do they predict how well a video does?

The main result

Finding 1 · Nothing predicts performance

To test prediction we standardised every metric within each quarter, so era effects can't leak in, and correlated it against performance. Here is every correlation we measured, against views adjusted for channel size:

Correlation between thumbnail properties and performance

Pearson r vs log(views per 1,000 subscribers), z-scored within quarter · n = 440

The shaded band is the region where a correlation is statistically indistinguishable from zero at this sample size (|r| < 0.094, p > 0.05). Five of seven metrics sit inside it. Of the two that don't, only colourfulness survives correction for testing seven hypotheses at once — and the next section shows that one is an artefact.

Against raw view counts, with no adjustment at all, the picture is even flatter — the largest correlation of the seven is r = 0.09, which is roughly 0.8% of the variance. Brightness, saturation, colourfulness, contrast, busyness: none of them move the needle.

Metric	r (vs views)	r (vs views/subs)	Verdict

The one effect that looked real

Finding 2 · And why it isn't

Colourfulness came back at r = −0.152 (p = 0.0014) — comfortably significant, and significant even after correcting for seven tests. Read naively it says something genuinely interesting: less colourful thumbnails overperform, the exact opposite of standard YouTube advice.

It is not true. The outcome variable — views per subscriber — is mechanically lower for large channels, because a channel with 10 million subscribers cannot get 50 views per subscriber. So any thumbnail property that correlates with channel size will produce a fake correlation with that outcome. We checked:

The confound

Colourfulness correlates with channel size at r = +0.133 (t = 2.80, p = 0.005). Bigger channels make more colourful thumbnails — they have designers, brand systems and budgets. Bigger channels also mechanically have lower views-per-subscriber. That chain alone predicts a spurious negative correlation of roughly the size we observed.

Once channel size is accounted for, the effect is not distinguishable from zero. What looked like a thumbnail finding was a production-budget finding.

This is the trap the whole category falls into, and it's worth stating plainly: polished thumbnails correlate with success because successful channels can afford polish, not because polish causes success. Correlational thumbnail advice almost never separates those two.

Did anything change over time?

Finding 3 · Barely

The second question was whether the look of a niche visibly drifts. Each panel below is one metric, quarter by quarter, on its own scale. r² is how much of the movement a straight-line trend explains.

Quarterly means, 2024-Q1 → 2026-Q3

Each panel on its own y-scale · 40 thumbnails per point

Edge density has the most consistent trend (r² = 0.42) — thumbnails in this niche got about 11% busier. Dark pixels moved further in absolute terms (+40%) but far less steadily (r² = 0.28). With eleven time points and seven metrics tested, neither survives correction for multiple comparisons. Treat both as hints, not findings.

So the honest answer to “is this niche's visual style evolving?” is: if it is, the effect is smaller than the noise between individual thumbnails in any single quarter. Within one quarter the spread on every metric is wider than the entire 2.5-year range of the quarterly averages.

What this does and doesn't mean

Limitations — read these

We are reporting a null result, and null results are easy to overclaim. So, precisely:

This study cannot tell you thumbnails don't matter

It tells you that these seven pixel statistics, in this one niche, don't predict performance. That is a much narrower claim.

The specific weaknesses, stated openly:

Range restriction. We sampled the top 40 per quarter, so every video in the set is already a winner. Correlations are attenuated when you only look at the top of the distribution. A sample including mediocre videos would be a stronger test.
Crude proxies. Skin-tone fraction is not face detection. Edge density is not text detection. The things that probably do drive clicks — who is in the picture, what the text says, what emotion is on the face, whether it matches the title — are semantic, and we measured none of them.
Views are not click-through rate. A thumbnail's actual job is CTR on the impressions it receives. We can only see public view counts, which are mostly determined by distribution, not by the image.
One niche, one language bias. “AI tools” is a chaotic, fast-churning category with a very mixed creator base. Results may differ elsewhere.

What we'd claim with confidence is narrower and still useful: if a tool scores your thumbnail on brightness, saturation, contrast or colourfulness and tells you that score predicts performance, it is not supported by this data. Those numbers are descriptive, not predictive, and should be sold as such.

Measure your own thumbnail

Live tool · runs entirely in your browser

Paste any YouTube URL or video ID. Your thumbnail is measured on the same seven properties and placed against the distribution of all 440 thumbnails in the study. Nothing is uploaded — the analysis happens on your device.

These percentiles are descriptive, not a score. Being at the 90th percentile for saturation is not “good” — this study found no evidence that it predicts anything. It tells you where you sit relative to what is currently ranking, which is useful for knowing whether you blend in or stand out. That is a different question from whether you'll win.

Data & reproduction

Method detail

Thumbnails fetched at maxresdefault (falling back to hqdefault), downscaled to 160×90 and measured in a canvas context. Colourfulness follows Hasler & Süsstrunk (2003). Skin detection uses the standard YCbCr bounding rule. Edge density is the mean Sobel gradient magnitude over the luminance plane. Video metadata — publish date, views, subscriber count — came from the vidIQ API, sampled by quarter and ordered by view count. All correlations are Pearson on within-quarter z-scores; significance is a two-tailed t-test with 438 degrees of freedom; the multiple-comparison threshold is Bonferroni at α = 0.05/7 = 0.0071. The analysis pipeline is the same code running in the tool above — you can read it by viewing the source of this page.

The full run was executed twice from a cold start and reproduced identically to the displayed precision.