A new study of thumbnail design across 1,043 YouTube channels with at least 100,000 subscribers found that the “shocked” or “surprised” expression, treated as standard advice in most thumbnail guides, appears on just 11% of the covers that show a face. “Neutral” expressions accounted for 35% and “happy” expressions 32%, according to the report from Thumbnail Bench, an AI thumbnail-generation company.

The study tagged 21 design features, including face count, expression, word count, color, and layout devices such as arrows and screenshots, across 11,223 long-form thumbnails pulled from the 15 most recent uploads of each channel. Shorts were excluded through three separate checks, reducing an initial pool of 16,499 feed items to 11,223 covers. The channels spanned 11 niches, sampled at up to 100 per niche.

Faces appear on 71% of the covers studied, with 46% showing a single person and 25% showing two or more. How often a face appears, and what it is doing, varies sharply by niche. Vlog, Fitness, and Finance channels put a face on close to nine in ten covers, while Gaming sits at 47% and Education at 57%. The shocked expression is similarly niche-dependent: it shows up on a quarter of Gaming faces but just 3% of Fitness faces, with Finance at 12% and Commentary at 16%.

Where a face does appear, it covers more than a third of the image height on 33% of covers, is medium-sized on 36%, and small on 31%. Eye contact with the camera runs to 49% overall, rising to 74% in Finance and 63% in Vlog.

Word Counts Undercut the ‘Three Words Maximum’ Guidance

Seventy-seven percent of covers carry designed text, and among those the median word count is four. Twenty-two percent use one or two words, 32% use three or four, and 27% run to seven or more, meaning the widely cited three-word cap holds on only about four in ten text-bearing covers. Twenty-three percent of all covers carry no text at all, a share driven heavily by Food (51%) and Gaming (33%), while Finance runs the opposite way at 6%.

Text skews white (68% of covers with words), set in capital letters (62%), and given an outline, shadow, or box (73%). Fifty-six percent of covers carry a warm color accent (red, orange, or yellow) over at least 15% of the frame; in Food that figure reaches 82%.

The study also checked how much of each cover’s text duplicates the video’s title. On 36% of covers, half or more of the words match the title. Fitness channels repeat the title most, at 52%, and Food channels least, at 23%.

Larger Channels Use Fewer Devices, Not More Faces

Splitting the sample into three subscriber bands, 100,000 to 500,000 (5,400 covers), 500,000 to 2 million (3,572 covers), and above 2 million (2,251 covers), the face rate holds nearly flat: 72%, 68%, and 74%. Other devices drop as channels grow. Channels above 2 million subscribers go text-free on 32% of covers versus 19% for the smallest band, use arrows on 9% of covers versus 15%, and use screenshots on 7% versus 15%.

Finance emerged as the most formulaic niche in the sample, with a face on 88% of covers, eye contact at 74%, text on 94%, and the highest arrow rate of any category studied. Food sits at the other end, with 51% of covers carrying no text and 82% relying on warm color.

No Single Feature Stood Out on Winning Covers

For 609 channels with at least eight long-form covers older than a week, the study compared each channel’s highest-viewed-per-day cover against its other covers on every tagged feature. The gaps were narrow across the board: a face appeared on 72% of the top-performing covers versus 70% of the rest, and most other features fell within two or three percentage points. The only differences to clear that noise were small: top covers repeated the video title less often (33% versus 35%) and carried a face slightly more often.

Thumbnail Bench notes that views per day is a proxy, not click-through rate, which is not public, and that the metric reflects a video’s topic, upload timing, and how the recommendation system handled it, in addition to the cover itself.

Thumbnails were tagged using Claude Opus 5, with accuracy checked through a pilot comparison; a Claude Sonnet 5 run was tested and rejected after it agreed on only 88% of fields.

Image source: Thumbnail Bench
The full report is available here

Subscribe to Our Newsletter

Check Out Our Podcast