Last month I pulled 30 of our own thumbnails and did something dumb but useful. I printed each one at postage stamp size, taped them to a wall, stood back six feet, and asked three people in the studio one question: where did your eye land first?
On 22 of the 30, the answer was the same. The eye. Not the face. Not the subject. The eye.
That gap between "put a face in it" and "put an eye in it" is most of what people get wrong about faces in thumbnails, and it is the reason a lot of face-forward thumbnails still underperform.
A face is not a magnet. A gaze is.
The standard advice is that human faces draw attention because we are wired for them. True, roughly. But the wiring is more specific than the advice suggests. We do not scan for faces as shapes. We scan for eyes, because eyes tell us what a person is looking at and how they feel about it. A face turned away, a face in shadow, a face too small to resolve eyes at all: those are just skin-toned blobs in the layout. They occupy space and return nothing.
I see this constantly in first drafts, including my own. A designer crops a wide shot down, the subject is now roughly a third of the frame, the face is technically visible, and the thumbnail feels fine on a 27 inch monitor. On a phone the eyes are eight pixels tall and the whole thing reads as a beige smear next to text.
The test I use now is blunt: if you cannot tell where the person is looking, the face is decoration.
Eyes should be big enough to have direction
There is a size threshold under which eyes stop functioning as eyes. In our own testing I have found it sits somewhere around 3 to 4 percent of thumbnail height for the eye region. Below that the pupil, the iris and the whites collapse into one dark mark and the viewer loses gaze direction entirely.
This is the practical reason so many creators crop tighter than feels comfortable. Not because close-ups are trendy. Because a close-up is the cheapest way to guarantee the eyes survive compression to a 168 pixel wide phone tile. If you have never checked what your work looks like at real size, this is the single fastest fix available to you, and I have written before about the mistake of designing at full size because it is the root cause of about half the thumbnails I have to redo.
Tighter crops also fix a second problem nobody mentions. When the eyes are large, you can direct them. A subject looking slightly off axis toward your text block pulls the viewer's attention there without a single arrow. When the eyes are tiny, you need arrows, and arrows cost you clutter.
Gaze direction is a free pointer, and it beats a drawn one
Here is my actual opinion, and people push back on it: a well aimed gaze outperforms a red circle in almost every case where both are options.
The reason is that gaze is read pre-attentively. You do not decide to follow someone's eyeline. It happens before conscious processing and it costs zero visual weight. A red circle has to be drawn, which means it adds a shape, a color, and an edge to a frame that is already fighting for room. It buys attention with clutter.
I am not saying circles and arrows are dead. On dense screenshot thumbnails, tech reviews, gameplay, anything where the point of interest is a small object in a busy field, a marker still does work that a gaze cannot. I went deeper on when arrows and circles still work and the honest summary is that they earn their place on complex frames and get in the way on simple ones.
If you can point with a person's eyes instead of a graphic, do that. Every shape you do not draw is contrast you get to spend somewhere else.
The sequencing matters too. I like to build the eyeline first, place the text where the eyeline lands, then check whether anything still needs a marker. Usually it does not. When you go the other way, text first and gaze second, you end up with a subject staring off the left edge while the payload sits on the right, and the thumbnail quietly fights itself.
Direct to camera versus off axis
Two modes, two jobs.
Direct to camera creates a social contact effect. The viewer feels addressed. It works well for opinion content, personal stories, confrontational titles, anything where the promise is "I am going to tell you something." It also holds up at small sizes better than off axis, because a symmetrical face reads faster.
Off axis creates a scene. The subject is looking at something and the viewer wants to know what. It works when there is a second element worth looking at: a product, a chart, a result, a reaction. It fails badly when there is nothing at the end of the eyeline, which is a surprisingly common mistake. A subject staring at empty gradient looks confused rather than intriguing.
On a finance channel around 80k subs we ran an A/B variant across a run of videos where the only change was direct to camera versus the subject looking down at a chart. The chart version won on the videos where the chart was the story. The direct version won on the videos where the host's take was the story. The delta was in the low single digits of relative CTR either way, which is honest and unglamorous: this is a matching problem, not a ranking problem. There is no universal winner, and anyone who tells you there is has run one test on one channel.
Expression carries more than the crop does
Once the eyes are large enough and pointed correctly, the next lever is expression, and it is a bigger lever than most people expect.
The failure mode I see most is the flat professional face. Neutral, composed, camera ready, completely uninformative. It tells the viewer nothing about the emotional register of the video. Compare that with an expression that has a clear read at a glance: skeptical, delighted, alarmed, deadpan. Any of those give the viewer a reason to guess what happens in the video.
The second failure mode is overshooting. The screaming open mouth face has been so overused in some niches that it now signals low quality to a chunk of the audience. In beauty and lifestyle work in particular I have watched restrained expression beat exaggerated expression, which is part of why our approach to beauty thumbnails and close ups leans on skin, light and eye contact rather than theatrics.
A useful middle setting: pick one emotion, push it about 70 percent of the way, and let the eyes carry it rather than the mouth. Eyebrows and eyelids do more per pixel than lips do.
Where the eye actually travels, and how to stop wasting the second stop
Attention on a thumbnail is not one landing. It is roughly two stops in the first second. The first stop is almost always the face if there is one, specifically the eyes. The second stop is wherever the composition sends you next.
Most thumbnails win or lose on the second stop. If it lands on your text and the text is legible, you have communicated. If it lands on background noise, a logo, or text that is too small, you have spent the viewer's entire attention budget on a hello with no follow up.
That is why legibility work is not a separate discipline from face work. They are the same problem. The reason your text is unreadable on mobile is usually that it was sized against a face that was itself too small, so everything scaled down together. Fix the crop and the text sizing tends to fix itself, or at least become obvious.
Font choice compounds this. Heavy geometric sans faces hold their shape next to a high contrast face better than anything decorative, and we keep a running set of rules in our thumbnail fonts guide for exactly this reason.
Shorts changed the geometry, not the principle
Vertical formats moved the safe zones and squeezed the usable area, but they did not change what the eye does. If anything, gaze matters more in a 9:16 frame because there is less room for markers and text, so pointing with a look is often the only pointing available.
The practical shift is vertical stacking. Eyes in the upper third, text below, nothing important near the bottom where the UI chrome sits. I covered more of what actually changed for shorts thumbnails elsewhere, but the short version is that faces got more important, not less, because there is no room for anything else.
What we actually do about it
Every thumbnail we make is custom per video. No templates, no swapping a face into a preset layout, because a layout that was built for a different eyeline will fight the new photo every time. We also test new styles on our own channels before they go anywhere near a client's, which is how the eye size threshold above came from our own failures rather than a blog post.
The pack is $250 for 8 custom thumbnails, about $31 each, one per video, with unlimited revisions and a dedicated designer. Title consultation is included, because the text and the gaze have to agree with each other and you cannot fix that after the title is locked. Turnaround runs 48 to 72 hours per thumbnail. If you want to test gaze direction properly, the A/B variant add-on is $15 and it is the cheapest real experiment available on this list. There is a title pack at $25 and rush at $20 if you need them.
If you want to see how this thinking shows up in finished work rather than in theory, our work is the honest version of this article. We never show client names, so what you get is niche and scale, not logos.
One more thing, since it comes up: if you are also sitting on long form interview footage, our sister studio handles video podcast editing and that is a different conversation entirely.
Faces are not a hack. They are a pointer with an emotion attached. If the eyes are too small to resolve, you have no pointer. If the eyeline aims at nothing, you have no destination. If the expression is neutral, you have no emotion.
Get those three right and you rarely need an arrow. Get them wrong and no amount of red circles will save the frame.
If you are staring at a thumbnail right now and it feels flat, crop in until the eyes scare you a little. That is usually the right size.