Should all your thumbnails look the same?

The advice contradicts itself

Two rules get repeated in the same breath. Be consistent, so a viewer recognises you in a feed. And make every thumbnail specific to its video, so it earns the click on its own. Taken literally they pull in opposite directions, and neither side usually says which parts of a thumbnail they mean.

That is answerable rather than arguable. The packs in this tool hold sixty thumbnails per niche drawn from a handful of channels each, so for thirty-three channels there are enough of their own thumbnails side by side to ask: what does this channel actually hold constant, and how does that compare with the variation across its whole niche?

What I measured

Thirty-three channels with at least eight thumbnails in the packs, twenty for most of them, 660 images in total. For every image, five numbers.

Two are the same measures as in the guide on standing out: mean relative luminance by the WCAG formula, and mean HSL saturation. The third is the dominant hue bucket, one of twelve 30° sectors, weighted by saturation.

The last two are proxies for composition. I split each frame into vertical thirds and recorded which third is brightest, and which has the most horizontal detail — the sum of luminance differences between neighbouring pixels. A face or a subject is usually both the brightest and the busiest region, so “which third wins” is a cheap stand-in for where the subject sits. It is a proxy, not a subject detector, and the guide treats it as one.

Then, for each channel, the spread of the first two measures and the share of its thumbnails agreeing on the last three — and the same statistics computed over its whole niche, sixty thumbnails from several channels, as the control.

What channels actually hold constant

Medians across the thirty-three channels, against medians across the eleven niches. Lower is more consistent for the spreads; higher is more consistent for the shares.

The ratios are the finding. If a channel style were mostly a fixed look, the first two rows would be far below 1 and they are not: brightness comes out at 0.94×, which is to say barely distinguishable from the variation across an entire niche.

Brightness is not a channel trait

The spread of brightness inside a channel is 0.106 against 0.113 across its niche. Twenty-one of the thirty-three channels are tighter than their own niche, which is close to what you would get from nothing in particular, and the range across channels is enormous: from 0.027 to 0.158, a factor of six.

That is worth sitting with, because “keep a consistent look” is usually implemented as a fixed exposure and a fixed background. What the measurement says is that the channels in these packs mostly do not do that. Brightness in a thumbnail tracks the content — a night scene, a bright studio, a white product on a table — and the channels leave it free to.

Palette is

Saturation is the first measure that separates a channel from its niche properly: 0.109 against 0.141, a spread 23% tighter. And the leading hue agrees on 70% of a channel's thumbnails against 60% within a niche.

The distribution is more useful than the median here. Seventeen of thirty-three channels have the same leading hue on at least 70% of their thumbnails, and eight are at 90% or above. That is a real house palette, and it is the cheapest identity to recognise at the size a thumbnail is delivered — 336 pixels wide on a phone-class screen, as measured in the guide on blurry thumbnails. Hue survives that reduction when nothing else does.

And so is where the subject sits

Both composition proxies come out at 55% within a channel against 48% within a niche, and this is where the largest number of channels beat their own niche: twenty-four of thirty-three on the brightest third.

Fifty-five percent sounds low for a “layout”, and it is worth reading carefully: it means the modal third wins about half the time, in a measure with only three possible answers. Seven channels are above 70%, and one — Homestead Roots — puts the brightest third in the same place on 94% of its thumbnails. A viewer scanning a column does not read a layout, but a repeated one means the eye lands in the same place twice, which is what recognition is at that size.

The two ends of the range

The extremes are more instructive than the medians, because they are two different strategies rather than one done well and one done badly.

Fix the axis that identifies you, free the one that describes the video

Andrei Jikh is a template in the strict sense: brightness spread of 0.027, 95% one hue, 75% one composition. Every thumbnail is the same object with different content in it, and it works because the niche it sits in is crowded and the recognition is worth more than the per-video fit.

Jeff Nippard is the more interesting case and, I would argue, the better default. His leading hue is the same on 100% of the thumbnails measured while his brightness spread, 0.156, is the widest of any channel here. The colour identifies the channel; the exposure describes the video. Both jobs get done, and neither constrains the other.

NikkieTutorials sits at the same brightness spread with hue at 70%, in the brightest niche measured — a row where more than half of everything is already light, per the niche measurements. In that context variation is not inconsistency so much as it is the only room left.

Which suggests a rule with numbers behind it: lock the palette, keep the subject in the same third, and let brightness follow whatever the video actually looks like. That is what the median channel in these packs does, and it is the opposite of a template that fixes exposure and lets colour wander.

Why this shape and not the other one

It follows from where a thumbnail is seen. A row in a feed is mostly other channels — your own videos rarely appear next to each other — so the consistency you are building is not read as a set, it is read one cell at a time against strangers. What that rewards is a signature recognisable in isolation, and hue plus position is exactly that, while brightness in isolation says nothing about who made it.

It also follows from the delivery size. At 336 pixels the things that survive are large areas of colour and gross placement, which is the same reason the floors in the colour guide and the guide on faces are set where they are. A house style made of small details is a house style nobody receives.

What this measurement is not

Thirty-three channels is a small sample, and they were selected for this tool's practice packs — popular channels in eleven niches — not sampled at random. The right reading is “this is what channels that are doing well look like”, not “this is what works”.

There are no view counts in the packs, so nothing here says a consistent palette earns clicks. What it establishes is a description: on which axes successful channels are consistent, and by how much relative to their own niche.

And the two composition measures are proxies. Brightest third and busiest third correlate with where a subject is, and they also fire on a bright sky or a busy background. They are informative in aggregate and unreliable on any single image, which is why the numbers here are shares across a channel and not verdicts about a thumbnail. It is also further than the platform goes: YouTube's thumbnail and title tips recommend consistency without saying which part to hold.

Doing this on your own channel

The whole measurement is a summary of something you can look at directly, and your own channel is the easiest case: export your last twenty video URLs into a CSV, load them as your competitor list, and you are looking at your own feed row. Two questions answer themselves. Do they share a palette, and does the eye land in the same third?

Then swap in your niche pack, which is the row you are actually in. That is what the preview tool is for, and it is the same check the guide on common mistakes keeps returning to: judge the row, not the image.

The questions everyone asks