Skip to content

Do YouTube Thumbnails with Larger Faces Get More Views?

Published
2026-06-15
Version
1.0
Data & code
Replication repository
https://github.com/HitFactorApp/youtube-thumbnail-face-size-study
Cite

“Put a face on it” has a louder cousin: fill the frame with it. Make the face big, the advice goes, and the views follow. Our earlier study tested whether to use a face at all (within a channel, a face is tied to slightly fewer views). This one tests the size half of the claim, on the videos that already have a face.

Comparing each channel’s faced videos against its own faced videos, holding age constant, a bigger face changes views by essentially nothing.

A bigger face −0.6% views

Doubling how much of the frame the face fills does essentially nothing to views.

We're 95% sure the real number is between −3.3% and +2.1% — a tiny range that sits well inside the 10% line we set as worth acting on. A confident near-zero, not a 'we couldn't tell.'

This is not an absence of evidence. With ~2,200 channels and ~68,000 videos the estimate is tight: the whole confidence range runs from about −3% to +2%, so any real size effect is small. The data say making the face bigger does not move the needle.

We measured every faced thumbnail’s largest face as a share of the frame, then asked how a channel’s own views move as its own face size moves, holding video age constant. Comparing a channel to itself cancels out channel size, niche, and audience, so the result cannot be “big channels just shoot bigger faces.”

The plan was written and publicly timestamped before any size-vs-views number was queried, and the headline was read once off a sealed set of channels we never touched while building the analysis.

Before the finding, one idea that makes this study make sense: a result can be real and still be too small to care about.

With tens of thousands of videos, almost anything you measure will be “statistically real” — that just means we’re confident it isn’t exactly zero. But “not zero” is a low bar. A change can be real and still be so tiny that no creator would ever notice it or change a thing because of it. Doctors hit this all the time: a drug can have a “real” effect that’s so small it doesn’t actually help the patient. Real on paper, useless in practice.

So we don’t ask “is there any effect?” We ask “is the effect big enough to act on?” To answer that, we drew a line before looking at any results: 10%. A change has to move views by more than 10% (for a doubling of the face’s size) to count as worth acting on. Ten percent is our honest call for the smallest change a team would actually do something about. Drawing the line in advance means we can’t slide it around later.

That’s the lens for the finding below: making the face bigger lands at essentially zero, and the whole confidence range sits well inside the 10% line — so this isn’t “we couldn’t tell,” it’s a confident “this doesn’t matter.”

A bigger face is tied to no change in views

Doubling the face’s share of the frame moves within-channel views by −0.6% (95% CI −3.3% to +2.1%). The separate set-aside group agreed at −0.2%. Confidently near zero.

Flat at every size

Split faced videos into size quartiles, from a median 2.7% of the frame up to 8.6%, and views are flat across all four. No size where bigger starts to pay, no sweet spot in the middle.

Not a face-count effect in disguise

Restrict to single-face thumbnails, so size can’t stand in for “more faces,” and the result is still near zero (+0.2%).

Not an artifact of the cutoff

Raise the bar for what counts as a face from 2% to 4% of the frame and the slope stays flat (+0.3%).

The headline and all three pre-registered checks sit on top of zero, their intervals straddling it, every point far inside the ±10% bar.

Views per doubling of face area (held-out channels)
-10 -5 0 +5 +10 Headline (all faced) Single-face only Stricter 4% floor % per doubling

Whiskers are the 95% interval, resampled across whole channels. The dashed ±10% framing is the practical bar set in advance; every estimate is nowhere near it.

With each channel and video age removed, the four within-channel size quartiles line up flat. It’s the same near-zero result shown as a shape: there’s no curve to find.

Residual views by within-channel face-size quartile (held-out channels)
-5 -2.5 0 +2.5 +5 Q1 · ~2.7% Q2 · ~3.6% Q3 · ~5.1% Q4 · ~8.6% relative log-views

Each quartile’s interval covers every other quartile’s value. Quartile labels show the median face size in that bin.

The data. Faced videos from English-language business and creator channels: public, organic, long-form, each channel with at least 20 qualifying videos. The headline is read off a held-out 20% of channels: 2,193 of them have at least five faced videos that vary in size, and 2,192 enter the slope (67,580 videos). Not a random slice of YouTube.

The comparison. Each channel against itself (channel fixed effects), holding video age constant, with face size and views both on a log scale, so the headline is one number: the percent change in views per doubling of face area. Because the whole set was measured on one clock, a video’s age is effectively when we captured its views; controlling for age removes that timing artifact.

Clustered inference. Videos in a channel are not independent, so every confidence interval resamples whole channels (10,000 times), re-centering within each draw, never resampling individual videos.

The face-size check. Face size is the detector’s box as a share of the frame. We confirmed the detector’s boxes stay tight and consistent across the size range with a stratified spot-check of 50 faces. Because the result is near zero, this matters less than it would for a positive finding: a size-dependent measurement error would manufacture a slope, and no slope is present.

The statistics. At this scale a p-value is near zero for any real effect, so we report the effect size and a channel-clustered interval against a ±10% practical bar set in advance. Everything here is a correlation; nothing was randomized.

  • It’s a correlation, not a cause. Face size travels with framing, topic, and the rest of a packaging choice we can’t fully separate. “Tied to,” never “causes.”
  • It’s a conditional question. We look only at videos that already have a face, which is itself a packaging choice related to views. We state the finding only among faced videos and bound that selection as small using the presence study; we don’t claim it’s zero.
  • These channels aren’t all of YouTube. Face-using channels that vary their face size, from our English-language business and creator funnel.
  • We measured views, not watch-time, captured once. Deleted and private videos are invisible.
  • The detector size check is the study’s softest step. A visual spot-check, not the numeric box-error bound originally planned (which needed labels that didn’t exist). It’s bounded by the detector’s 0.994 presence recall and matters less for a near-zero result.

For brand and marketing teams deciding how much to invest in the face on a thumbnail:

  • Don’t obsess over face size. The data doesn’t support cropping tighter or shooting a bigger face to move views. The effect is a precise zero, not “we couldn’t tell.”
  • Pick what fits the scene. If you’re going to put a face on the thumbnail, choose the framing that suits the shot. Since size doesn’t move views, it’s a creative call, not a lever to pull.
  • Test it on your own channel. The result is precise but correlational, so the only way to know for sure for your audience is to run your own A/B test.
  • Data: Hitfactor’s dataset of YouTube channels and videos. Face detection by a standard open model.
  • Conflict of interest: Hitfactor builds tools for video teams. We wrote and timestamped the analysis before seeing the result so the outcome couldn’t bend the method.
  • Replicate it: the binned tables and a de-identified row dataset, plus a notebook that reproduces the headline from that dataset alone, ship in the study repo.
  • Cite as: Hitfactor (2026). A larger thumbnail face changes views by essentially zero (−0.6%): a within-channel dose-response study across 2,192 YouTube channels. hitfactorapp.com/labs/2026-thumbnails-with-larger-faces/