YouTube Thumbnail Facial Expressions: Which Faces Still Read at 168 Pixels Wide

Almost every guide to YouTube thumbnail facial expressions is a vibe chart. Shocked for reveals, surprised for discoveries, smiling for tutorials, confused for mysteries. The charts are not exactly wrong, and they are completely untestable, which is why there are so many of them and why they all agree.

None of them asks the question that comes first. Before an expression can mean anything it has to be recognised — in an image 168 pixels wide, usually sitting in the corner of someone's vision while they look at something else. That is a measurable property of a face, it varies enormously between expressions, and vision scientists have measured it. Some expressions are built from coarse information that survives being shrunk to almost nothing. Others depend on fine detail around the eyes and mouth that is gone by the time the image reaches the feed.

First, the uncomfortable data: does a face help at all?

The figure you have read a dozen times is that thumbnails with faces get around 20% more clicks. Chase it and it does not resolve into a study. It resolves into other blog posts.

The most careful public analysis covers 1.25 million videos from 16,737 channels, and its real finding is methodological. Compare videos across different channels and faces look mildly positive, about 3% more views. Keep the comparison inside one channel — same creator, same audience — and a face is associated with roughly 6% fewer views. The cross-channel number is confounded in an obvious way: larger channels use faces more and get more views anyway, so part of the apparent benefit of faces is the benefit of being large. The same analysis found no measurable gain from a second or third face.

A separate analysis of roughly 300,000 viral videos from 2025, reported by Search Engine Journal, landed near parity, with variation by segment: a small benefit for larger channels, a clearer one in finance, and multiple faces outperforming single ones — the opposite of the within-channel result. Observational analyses disagree like this on a platform where nothing is held constant.

Against that, vidIQ's study of 500 breakout thumbnails found a human face was the commonest trait, and its share rose with performance: 69% across all 500, 75% of the top 100, 80% of the top 50. A real pattern, but a descriptive one — and a trait shared by 69% of a sample cannot differentiate much inside it.

The honest conclusion is the useful one. A face is not a lever. Putting one in frame buys nothing automatically, which is why so many thumbnails contain a face doing nothing — present, lit, centred, communicating no information. What a face carries is an expression, and an expression can carry the thing a thumbnail is short of: the state of the story.

Why thumbnail size, not taste, decides which expression works

A 2009 paper in Psychological Science by Fraser Smith and Philippe Schyns — "Smile through your fear and sadness" — asked which expressions survive distance. The method was to degrade faces the way distance degrades them, downsampling from 240 × 160 pixels to as low as 15 × 10, and measure which emotions people could still identify. Happiness and surprise came out as the two best recognised from a distance, because what distinguishes them lives in several low spatial-frequency bands: the coarse, blurry layer of an image that shrinking preserves. Fear and anger were recognised poorly across the range. The authors read an evolutionary logic into it — the two signals that work at a distance are the two you would want to send at a distance, while the rest were built for conversation range.

Nine years later a PLOS ONE paper by Smith and Stéphanie Rossit attacked it from the other direction, testing identification of expressions in peripheral vision. Fourteen participants saw faces centrally and then 15 and 30 degrees to either side. Recognition fell with eccentricity, and the expressions holding up best out there were again happiness and surprise. Summaries differ on mere detection — noticing a face is expressive at all — so treat identification as the robust part.

Now map that onto where your thumbnail appears. In a sidebar or a mobile list it renders close to 168 × 94, and it is almost never the thing being looked at: the viewer's fovea is elsewhere in the grid while peripheral vision decides whether your image deserves a fixation. Small and off-centre is the normal viewing condition, and it is exactly the condition under which the lab work says most expressions stop being legible.

This also explains a pattern creators notice without being able to justify. Theatrical expressions are not merely an excess of the MrBeast era; exaggeration is a rational response to a low-resolution channel. A wider mouth and higher brows push more of the expression into coarse shape rather than fine detail. The cost is everything else exaggeration does to how you are perceived — but the mechanism is real, and it is why the understated expression that looks sophisticated on your monitor dies in the feed.

What the expression is actually for

Thumbnail expressions are usually described as conveying emotion. That framing produces bad thumbnails, because emotion alone is not information: horror, delight and shock are available on demand and mean nothing by themselves.

The useful framing is narrative position: the face shows where its owner stands relative to an outcome the viewer cannot see yet. Mid-disaster is a position. About to find out is a position. Just found out, and it was worse than expected, is a position. Each implies a story shape without resolving it, which is the thumbnail's job — open the loop, leave it open, let the title and the first ten seconds close it.

Read that way the standard chart stops being arbitrary. A shocked face suits a reveal not because shock is clickable but because it puts the creator one step ahead of the viewer and promises the viewer will catch up. The same face on a spreadsheet tutorial works for nobody, because there is no outcome for it to be positioned against. So cover the text: if a stranger cannot tell from the face alone whether something good or bad has happened or is about to, the face is occupying the most valuable space in the frame to say nothing. Our rules for high-CTR thumbnails make the same point from the composition side — one focal point, carrying one claim.

The expression table

This combines the legibility findings above with what each expression can plausibly signal. The legibility column is grounded in research; the signal column is judgement, and should be read as a hypothesis to test rather than a law.

Expression	Legibility when small and off-centre	What it can signal	Common failure
Open-mouthed shock	Strong — mouth aperture is coarse shape and survives heavy downsampling	An outcome has landed, and it was extreme	Saturation: if every neighbour is shocked, shock is the baseline, not a signal
Surprise (raised brows, wide eyes)	Strong — one of the two best recognised at distance and in the periphery	Novelty, a discovery, an unexpected number	Reads as weak shock if the brow movement is small
Genuine smile	Strong — the other best recognised at distance	Warmth, a good outcome, safe-to-follow guidance	Posed smiles read as stock photography (see below)
Fear	Weak — poorly recognised across viewing distances in the 2009 downsampling work	Stakes, danger, a situation out of control	Collapses into generic surprise or a strained grimace
Anger	Weak — also poorly recognised across the range tested	Conflict, a callout, an injustice	Depends on brow and mouth detail the small render destroys
Disgust	Moderate at best — built from nose and upper-lip detail	A verdict delivered, usually on a product or claim	Misread as confusion, and the likeliest to make viewers dislike you
Scepticism (one brow, asymmetry)	Weak — asymmetry is the first information shrinking destroys	A claim being tested, a review with a view	Looks like a neutral face at 168 pixels
Concentration or effort	Moderate, if the body sells it too	Difficulty, a feat in progress, craft	The face alone cannot separate effort from discomfort
Deliberate deadpan	High as a shape — legible because there is nothing to resolve	Dry humour, authority, contrast against a chaotic frame	Only works when something else in the frame does the work

The pattern is consistent. What survives is large, low-frequency movement: a wide mouth, lifted brows, raised cheeks. What fails is the subtle, socially sophisticated end of the range — which is unfortunate, because those are the expressions that look best at full size.

What creators actually use is a different question

It is worth knowing the base rate. A Kapwing analysis of top-performing thumbnails published in September 2024 classified the expressions it found and reported surprised and happy at roughly 27% each as the two commonest; its authors noted they had expected angry or sad faces to do better. Read it for what it is — a census of what creators choose, not a measurement of what works.

If anything it argues against imitation. Half the competitive set is drawn from two expressions, which makes those two the wallpaper your thumbnail is trying to stand out from.

The eyes: direct gaze and gaze cueing are different tools

Where the subject looks is a separate decision from what their face does, and the research is unusually clear about the trade-off.

Direct gaze — eyes into the lens — captures and holds attention. Faces with direct gaze presented in peripheral locations draw attention more than faces with averted gaze or closed eyes, and direct gaze holds attention once the eyes arrive. Newborns prefer faces that look back at them, which suggests the sensitivity is not learned. The mechanism is still argued over: a 2021 study in Attention, Perception & Psychophysics concluded the capture depends on surrounding facial context rather than the eyes alone, and a 2015 study found eye visibility neither necessary nor sufficient. What survives the disagreement is the effect.

Averted gaze does something else. In the gaze-cueing paradigm established by Friesen and Kingstone's 1998 paper in Psychonomic Bulletin & Review, a face with averted eyes appears centrally, then a target appears left or right. People respond faster to targets where the face was looking — even though gaze does not predict the target's location, and even when they know it does not. The shift is reflexive, and showed up both at very short delays and at delays of several hundred milliseconds.

So there is a genuine either/or:

• Direct gaze when the face is the subject. Reactions, commentary, personal stories — anything where your response to something is the product.
• Averted gaze when something else is the subject. The repaired engine, the burnt cake, the chart. Pointed gaze is a cleaner redirect than an arrow, because it also tells the viewer how to feel about what it points at.
• Not both. A face looking at the lens while gesturing at an object asks the viewer to resolve two instructions in the fraction of a second they have allocated. One loses, and you do not choose which.

The smile problem: Duchenne markers at 168 pixels

Smiles are the most-used thumbnail expression and the most frequently botched, for a reason worth knowing by name. A smile recruiting only the mouth is driven by the zygomaticus major. A smile that also engages the orbicularis oculi — the ring of muscle around the eye socket — raises the cheeks, narrows the eye aperture and crinkles the outer corners. That second component is the Duchenne marker, long treated as what separates felt smiles from posed ones.

That claim has weakened, which is good news for anyone in front of a camera. Work led by Sarah Gunnery found most people can produce the marker on request: around 71% of participants successfully imitated a Duchenne smile and 69% a non-Duchenne one, and the authors suggest even that understates the ability, since nobody was coached. The marker does not certify sincerity.

What it still does is change how a smile is read. A 2020 study in Frontiers in Psychology used botulinum toxin to block the lateral orbicularis oculi and found the resulting non-Duchenne smiles were rated less happy, less genuine, less felt and less spontaneous — a subtle effect, in the expected direction. Some researchers argue the effect partly reflects smile intensity rather than authenticity, since a bigger smile recruits more muscle. For thumbnails both explanations point at the same instruction.

And here the resolution problem returns. At 168 pixels the crinkles are gone; fine detail is the first thing a small render discards. What survives is the coarse geometry the same muscle produces: cheeks sitting higher, eyes narrower, the whole eye region reshaped. "Smile with your eyes" is therefore not a mood note but a description of two changes you can verify in a still — are the cheeks up, and has the eye aperture narrowed. A mouth-only smile at thumbnail size does not read as insincere so much as it reads as nothing, which is worse.

Expressions act on the viewer before the viewer decides anything

Ulf Dimberg's work with facial electromyography established that looking at an expression produces a matching movement in the viewer's own face. A 1982 study in Psychophysiology found happy faces increased activity in the zygomatic region while angry faces increased activity in the corrugator region, the muscle that draws the brows together. A 1998 follow-up found those distinct responses were detectable after only 300 to 400 milliseconds of exposure. A 2000 paper in Psychological Science with Monika Thunberg and Kurt Elmehed used backward masking to present faces too briefly to be consciously seen, and the matching responses still appeared. Later work in which participants were told not to react found they went in the expected direction anyway.

The timings matter. The decision to scroll past happens in roughly the same window as that mimicry, so by the time a viewer reads your title the expression has already nudged their own face into a register. A furious face does not merely communicate conflict; it recruits the brow.

Be careful where that chain ends. No study connects thumbnail mimicry to clicks. The research supports the mechanism, not the outcome, and that gap is where most thumbnail advice fabricates its confidence — our post on the psychology of clickbait thumbnails works through more of the boundary. The modest version is enough for the decision in front of you: the expression you choose is the mood your title gets read in, so choose one the video can pay off.

The shocked-face question, answered properly

Every discussion of thumbnail expressions arrives eventually at the open-mouthed shocked face and whether it is finished. The usual citation is MrBeast, who said in 2023 that his watch time improved after moving from the wide-open-mouth look to closed-mouth expressions. The remark spread widely, and other creators have offered readings of it, MKBHD among those suggesting viewers who click a shocked-face thumbnail may feel they have been baited.

Take it seriously, but notice what it is: one creator's account of his own private metrics, on the largest channel on the platform, with no published numbers. Our breakdown of MrBeast's thumbnail style covers what does and does not transfer from that channel.

The argument that does generalise is saturation, and it is not about taste. When a format works it spreads, and as it spreads it stops differentiating — not because audiences grew sophisticated, but because the contrast that made it visible is gone. That applied to red arrows and yellow outlines before shocked faces, and it will apply to whatever replaces them.

Which leaves the only real method: test it on your channel. YouTube's Test & Compare runs up to three thumbnails against each other on one video and decides by watch-time share rather than raw clicks — the right measure here, because the whole charge against an exaggerated expression is that it buys a click the video cannot honour. An expression that wins on clicks and loses on watch time is precisely the failure the tool exposes. Expression is also one of the few thumbnail variables you can hold everything else constant against: same crop, same light, same text, three frames from one shoot.

How to shoot expressions you can actually use

Expressions are not photographs, they are moments, and moments resist posing. The workaround most professionals converge on is to stop trying.

1. Shoot video, not stills. Record ten or fifteen seconds moving through the expression — neutral into surprise into shock and back — then pull frames. A sequence contains transitional faces no posing produces, and the best thumbnail frame is usually a transitional one. Our video frame grabber exists for this step.
2. Light the eye region first. The eyes carry a disproportionate share of the signal and are the easiest part to lose. Overhead lighting is the main offender: the brow casts shadow into the sockets and produces no catchlight unless the subject tilts their face up. A frontal or 45-degree source slightly above eye level, or a reflector bouncing light back up, puts a visible highlight in the eye and keeps the socket open. Larger sources make larger catchlights.
3. Frame tighter than feels right. A face that reads comfortably on your monitor is often too small at 168 pixels. Crop so the eyes and mouth — the two regions people rely on to judge expression — take a meaningful share of the frame.

Our guide to taking photos for thumbnails covers the camera side in more depth, including how to get usable results from a phone.

Generated faces, and the policy line

A generated or composited face is a legitimate option when the expression you need is one you cannot produce on demand. Two constraints apply. The first is YouTube's thumbnails policy, which is narrower than creators fear and sharper than they assume. Thumbnails breaching the Community Guidelines are not allowed; the listed prohibitions include violent imagery intended to shock or disgust; and a thumbnail must not mislead viewers into thinking they are about to see something not in the video. Enforcement can mean removal of the thumbnail and a strike, though a first offence may draw only a warning. Nothing there restricts expressive faces as such. What it restricts is a face promising an event the video does not contain.

The second is that the expression has to survive the render. Generated faces fail at thumbnail size for the same reason posed smiles do: the coarse geometry is approximately right while the eye region is subtly wrong, and the eye region is where the signal lives. That is the part of the job Thumblore is built around — generating a face with a specific expression at the resolution YouTube serves, then letting you check it at that size rather than the size you designed it. For channels whose thumbnails live on reaction shots, our reaction thumbnail generator is tuned for this exact problem.

Frequently asked questions

What is the best facial expression for a YouTube thumbnail?

There is no single best expression, but there is a defensible default: a large, clearly-read surprise or a genuine smile, because those are the two expressions research repeatedly finds survive shrinking and peripheral viewing. Beyond that, the right expression states where you stand relative to the outcome the video delivers. An expression that reads clearly and promises nothing is no better than no face at all.

Do shocked-face thumbnails still work?

They still read, which is not the same as still working. An open mouth is highly legible at small sizes, so it continues to do its mechanical job. The problem is saturation: when a large share of surrounding thumbnails use the same expression, it stops distinguishing you. Test it against a closed-mouth alternative and judge by watch-time share, not clicks.

Should you smile in a YouTube thumbnail?

A smile is among the best-recognised expressions at small sizes, so yes, for content where warmth or a positive outcome is the promise. The requirement is that it engages the muscle around the eye, raising the cheeks and narrowing the aperture. A mouth-only smile reads as empty at thumbnail size, because the mouth is doing all the work, and a mouth doing all the work is what stock photography looks like.

Do you have to look at the camera?

No. Direct gaze captures and holds attention, so use it when your reaction is the subject. Averted gaze reflexively pulls attention in the direction you are looking, so use it when an object in the frame is the subject. Choose one job for the eyes; a frame attempting both resolves into neither.

How big should the face be in the thumbnail?

Big enough that the eyes and mouth are individually legible at 168 × 94 pixels, which is usually larger than designers choose on a full-size canvas. Crop until the expression survives the shrink test, then stop — a face filling the frame leaves no room for the object, text or context that gives the expression something to be about.

Do thumbnails with faces get more views?

Not on the evidence available. The cleanest public comparison, which held the channel constant, found a face associated with roughly 6% fewer views, and cross-channel comparisons that find a benefit are confounded by channel size. Faces are common among top performers, but common enough that their presence explains little. Treat a face as a vehicle for information, not an improvement in itself.

The bottom line

The expression debate is usually conducted as a matter of taste, one camp defending theatrical faces as what the platform rewards and the other treating restraint as a mark of quality. Both argue a layer above the one that decides the outcome. A thumbnail is seen small and in the corner of someone's eye, and under those conditions most of the human expressive range is not available to you. Happiness and surprise travel. Subtlety does not. That is a property of the image, not a preference.

What you do with the expressions that travel is the part no research can decide for you. Place the face relative to an outcome. Give the eyes one job. Check that the cheeks moved, not just the mouth. Then run the three-variant test, because the only data describing your audience comes from your audience.

Thumblore handles the production half of that loop: generating the specific expression a video needs, at the size YouTube serves it, so testing an expression costs minutes rather than a reshoot. For the surrounding craft, our posts on what eye-tracking research does and does not show and thumbnail composition cover the frame the face sits in.