What makes a good YouTube thumbnail? We analyzed 507,478 videos

Thumbnail advice is usually one successful creator generalizing from their own channel. We wanted the version with a denominator. Eden indexes YouTube alongside five other platforms for its Discover research feed, which scores every video against its own channel's baseline — so we can compare the thumbnails of videos that broke out against the thumbnails of videos that flopped, from the same channels, at scale. Here is everything we found, including the parts that undercut the genre.

Key findings

• 84.2% of breakout thumbnails contain a human face — but so do 82.2% of underperformer thumbnails. Faces are the norm, not an edge.
• 86.6% of breakout thumbnails carry overlaid text. Going text-free is 25% more common among underperformers.
• The text sweet spot is 4 to 6 words: 30.6% of breakout thumbnails sit there, versus 26.5% of underperformers — the largest single gap we measured.
• Thumbnail text averages about 7 words, not the 3 that most advice recommends.
• 62% of breakout thumbnail text is in all caps.
• Only 6% of thumbnails repeat the video title verbatim, and 39% share no words with the title at all.
• The shocked face is a big-channel style: 11.7% of breakout thumbnails on 100k+ subscriber channels, 5.6% below 100k. Smiling thumbnails slightly underperform everywhere.
• Brightness, saturation, contrast, and warm colors showed no measurable difference between winners and losers.
• The biggest packaging levers we found live outside the thumbnail: videos over 40 minutes break out at 41.1% versus 26.9% for videos under 5 minutes.

A breakout, throughout this study, is a video that earned at least 2x its own channel's baseline. The full method is below.

How we ran it

• Corpus: 507,478 long-form videos (Shorts excluded) from 8,691 channels tracked in Eden's index, published between August 2025 and August 2026, each at least 7 days old at measurement so views had settled.
• Metric: each video's performance relative to its own channel's recent median, adjusted for the channel's audience size at publish time — the same outlier score Eden's Discover feed uses. A breakout beat its channel's baseline by at least 2x (33.7% of the corpus). An underperformer, for the thumbnail comparison, performed at or below its channel's baseline (42.1% of the corpus). Measuring against each channel's own baseline means a 900-subscriber channel's hit counts the same as a 9-million-subscriber channel's.
• Thumbnail analysis: we sampled 1,500 breakout and 1,500 underperformer thumbnails at random (capped at 3 per channel per group, so no single channel's style dominates), fetched the actual images, and extracted a fixed feature checklist per thumbnail — faces, expression, eye contact, text word count and casing, background type, graphics — with a vision model, spot-checked by hand against the images. Brightness, contrast, saturation, and hue were computed directly from the pixels. 2,960 thumbnails survived fetching and analysis (1,486 breakout, 1,474 underperformer).
• Reading the tables: with ~1,480 thumbnails per group, differences under about 2.5 percentage points are within noise. We say so wherever that matters.

Faces: the norm, not the edge

Feature	Breakout	Underperformer
Any human face	84.2%	82.2%
Exactly one face	55.6%	54.1%
Face fills over half the frame	15.8%	14.9%
Eye contact with the camera	32.4%	29.2%
Expression: neutral or serious	51.1%	47.4%
Expression: smiling	21.2%	23.4%
Expression: shocked	8.5%	7.7%

The most-repeated thumbnail rule — put a face on it — is what everyone already does. Faces appear on more than 8 in 10 thumbnails in both groups, and the 2-point gap between winners and losers is within noise. A face won't set you apart; it's the uniform.

Inside that uniform, two small edges are worth knowing. Eye contact leans winning (32.4% vs 29.2%). And expression matters in the opposite direction from what you'd guess: neutral or serious faces lead breakouts at 51.1%, while smiling thumbnails are actually more common among underperformers (23.4%) than breakouts (21.2%). The friendly smile reads as wallpaper; the face that looks like it's about to tell you something difficult does slightly better.

The shocked face deserves its own line, because it's the most imitated style on the platform. Overall it appears on just 8.5% of breakout thumbnails — a minority even among winners. Split by channel size, it's revealed as a big-channel habit: 11.7% of breakouts on channels over 100k subscribers, 5.6% below. Big channels can spend an established relationship on a screaming face; for a channel nobody recognizes yet, the same face is a stranger yelling.

Text: the clearest real edge in the study

Feature	Breakout	Underperformer
Any overlaid text	86.6%	83.2%
No text at all	13.4%	16.8%
1 to 3 words	26.4%	25.6%
4 to 6 words	30.6%	26.5%
7 or more words	29.6%	31.1%
All caps (of text-bearing thumbnails)	62.3%	59.7%
Text repeats the title verbatim	6.3%	6.8%
Text shares no words with the title	39.2%	39.1%

Text is where the winner-versus-loser gap is most consistent. Text-free thumbnails are 25% more common among underperformers, and the 4-to-6-word band is the single largest gap we measured (30.6% vs 26.5%). Seven-plus words tips slightly losing. So the working range is wider than the "3 words max" rule — text-bearing breakout thumbnails average 6.8 words — but there is a ceiling, and it arrives around word seven.

Two more numbers worth stealing. First, all caps is the house style of the platform: 62% of breakout thumbnail text. Second, almost nobody repeats their title on the thumbnail — 6.3% of breakouts — and 4 in 10 thumbnails share no words with the title at all. The thumbnail text and the title are two different lines of copy doing two different jobs: one names the video, the other adds tension to the image.

Color and brightness: nothing there

Pixel measure (0–100)	Breakout	Underperformer
Brightness (mean luminance)	40.9	40.5
Contrast (luminance spread)	23.5	23.1
Saturation	35.2	34.7
Warm-color pixel share	31.1	30.5

We computed these directly from the pixels, so there's no model judgment involved, and the answer is flat: breakout thumbnails are not measurably brighter, punchier, or warmer than underperformers. "Saturate it more" is the cheapest advice in every thumbnail tutorial, and in half a million videos' worth of outcomes it buys nothing detectable. Background type tells the same story — real photographed scenes (52% both groups), collages (~24%), and illustrations (~8%) split almost identically. The one background that leans losing is the screenshot: 2.6% of breakouts versus 3.8% of underperformers, though the sample there is small. Drawn-on graphics — arrows, circles, outlines — carry a small real edge (41.8% vs 38.1%).

The finding that matters more: the thumbnail is packaging, not strategy

If winning and losing thumbnails look this similar, what actually separates them? The concept the thumbnail is wrapping. Two cuts from the full 507,478-video corpus make the point:

Video length	Videos	Breakout rate
Under 5 min	43,259	26.9%
5 to 10 min	107,972	30.2%
10 to 20 min	162,523	33.0%
20 to 40 min	106,666	34.8%
Over 40 min	86,470	41.1%

Title length	Videos	Breakout rate
Under 30 chars	29,108	26.1%
30 to 44	90,057	30.8%
45 to 59	138,614	33.9%
60 to 74	114,309	34.8%
Over 75 chars	135,390	35.9%

The spread between a short video and a long one (14 points), or a terse title and a specific one (10 points), dwarfs any thumbnail-feature gap we found (4 points at best). These are observational numbers — long videos are also more ambitious videos — but the direction is hard to ignore: the video's substance and how specifically the title sells it move breakout odds far more than any styling choice inside the 1280x720 rectangle.

How to act on this

1. Meet the table stakes without agonizing: a clear subject (a face if the video is you), 4 to 6 words of text in caps that don't repeat the title. That's the whole consensus formula; execution beyond it measurably doesn't separate winners from losers.
2. Skip the shocked face until roughly 100k subscribers. Below that, the data says a neutral or serious face with eye contact is the better bet.
3. Reinvest the hours you were spending in the thumbnail editor into the two levers with bigger measured spreads: a more specific title and a more ambitious video.
4. Study the breakouts in your own niche instead of global averages. Eden's chat can pull the top-performing thumbnails for any creator or topic from this same index, reverse-engineer the formula — palette, composition, text treatment — and generate thumbnails in that style with your face and your words. The Discover feed runs the same beat-their-own-baseline analysis continuously, per creator, and finding what works in your niche shows the workflow.

Frequently asked questions

Do faces in thumbnails get more views?

Faces are on 84.2% of breakout thumbnails — but on 82.2% of underperformers too, so a face by itself is not an advantage; it's the platform norm. The face details that lean winning are eye contact with the camera (32.4% of breakouts vs 29.2% of underperformers) and a neutral or serious expression rather than a smile.

How many words should a YouTube thumbnail have?

Four to six is the strongest band in our data: 30.6% of breakout thumbnails versus 26.5% of underperformers. Text-bearing breakout thumbnails average 6.8 words, and seven or more words tips slightly negative. Having no text at all is the worse mistake — text-free thumbnails are 25% more common among underperformers.

Do shocked-face thumbnails still work?

They're a minority style even among winners: 8.5% of breakout thumbnails overall. They're twice as common on channels over 100k subscribers (11.7% of breakouts) as under (5.6%), which suggests the style spends recognition the viewer already has. If you're small, the data doesn't support copying it.

Should the thumbnail text repeat the video title?

No — and almost nobody does it. Only 6.3% of breakout thumbnails repeat the title verbatim, and 39% share no words with the title at all. Treat the thumbnail text as a second, complementary line of copy: the title names the video, the thumbnail adds the tension.

Do brighter, more saturated thumbnails perform better?

We measured brightness, contrast, saturation, and warm-color share directly from the pixels of all 2,960 thumbnails, and the two groups are statistically identical on every one. Legibility at small sizes still matters, but "make it pop" is not a measurable edge.

Where does this data come from?

From Eden's own social index, the same corpus that powers its research features across YouTube, X, LinkedIn, Instagram, TikTok, and Substack. For this study we took every indexed long-form YouTube video from the 12 months to August 2026 that was at least a week old — 507,478 videos from 8,691 channels — scored each against its own channel's baseline, and visually analyzed 2,960 of their thumbnails. No surveys, no vendor data, no single-channel anecdotes. Our Substack timing study runs on the same index.

We built this study because thumbnail advice is almost always one creator's experience wearing a lab coat. If you want this analysis running on your own niche — including studying the top thumbnails in it and generating yours in the same formula — Eden's free plan includes 100 saved items and 50 AI credits with no card required. Start free.