What Eye-Tracking Research Actually Says About YouTube Thumbnails

Almost every thumbnail guide on the internet contains a sentence that begins "eye-tracking studies show". Almost none of them names the study. The claim that follows is usually something comfortable — put a face in it, use three words, high contrast wins — and the citation, when you chase it, turns out to be another blog post citing a third blog post.

The underlying literature does exist. It is decades deep, mostly published in vision science rather than marketing, and it is considerably more interesting than the version that gets repeated. It also says less than the blogs claim and more than you would expect: it supplies a mechanism for several of the standard thumbnail rules, quietly withdraws support from a couple of others, and explains why the thing you cannot shake — the sense that a thumbnail is working before you can articulate why — is an accurate description of how the visual system actually behaves.

This post walks through what the research measures, what it has found, where it stops being a guide to YouTube specifically, and how to turn the parts that hold into decisions you make in front of an image.

What eye-tracking actually measures

An eye tracker records two things: fixations, the 200-to-300-millisecond pauses where the eye is nearly still and the high-resolution centre of your retina is pointed at something, and saccades, the fast jumps between them. Vision only works properly during fixations. Everything outside the fovea — which covers about the width of your thumbnail at arm's length — arrives at progressively lower resolution, lower colour fidelity and lower reliability the further out it falls.

This is the first thing the popular version gets wrong. A heat map is not a map of interest. It is a map of where the eye had to point in order to resolve detail. Something can be seen perfectly well in peripheral vision and never earn a fixation, because the periphery already answered the question. A large red shape in the corner of a thumbnail might be fully registered and never looked at. The absence of a hotspot is not proof that nobody saw it.

The second limitation is the task. Fixation patterns change dramatically depending on what the viewer has been asked to do — free viewing of a photograph produces a different scan path from searching that same photograph for a specific object. Most of the foundational studies use one of those two tasks on static images shown alone on a screen. Scrolling a feed of twenty competing thumbnails on a phone, with a personalised set of candidates and a subscription history, is a third task that the classic literature mostly does not cover.

So the honest frame is this: the research tells you what the visual system does with images in general, reliably and in detail. It does not tell you what your audience will do with your thumbnail. Those are different claims, and keeping them apart is the whole discipline.

The one study that put a YouTube page in front of people

There is very little published eye-tracking work on YouTube thumbnails specifically — if somebody quotes you a large dataset of thumbnail gaze patterns, ask for the paper. The closest thing to a direct measurement comes from an unexpected source: a 2023 study of trainee physics teachers choosing instructional videos, which tracked where they looked while deciding what to watch.

The result is worth the attention of anybody who designs packaging. Averaged across participants, the thumbnail accounted for 35.39% of total fixation duration — the only region of interest to clear 10%. The title took 9.58%. The channel name took 6.17%. Everything else, the view count, the upload date, the description snippet, drew comparatively little.

Treat the exact numbers as indicative rather than universal. The sample was narrow and specific: trainee teachers, an explicitly educational selection task, a constructed results page. Nothing about that generalises cleanly to someone half-watching a home feed on a television. But the shape of the finding matches what every creator's analytics already imply, and it is the rare case where the shape has been measured rather than asserted: the image is where the deciding happens, and the title is a confirming read, not a parallel one.

That has a practical consequence most packaging advice inverts. If the title mostly confirms a judgement the image has already triggered, then a title that repeats the thumbnail wastes the confirmation step. The two should answer different halves of one question, which is the argument made at length in our guide to writing titles that get clicks.

Faces and text are not stylistic choices

The most load-bearing finding in this literature, for thumbnails, is a 2009 paper in Journal of Vision by Moran Cerf, E. Paxon Frady and Christof Koch, Faces and text attract gaze independent of the task. The researchers showed viewers natural scenes containing faces, text and — as a control object — mobile phones, under both free viewing and visual search.

Under free viewing, participants fixated faces 16.6 times more than control regions matched for size and position, and text 11.1 times more. Two details matter more than the raw multiples. First, text came close to faces; the written word is nearly as strong a gaze magnet as a human face, which is not how most people intuit it. Second, the effect persisted when it was costly: the authors report that it is difficult to avoid looking at faces and text even when doing so carries a penalty in the task. These are not preferences. They behave like reflexes.

The paper's second half is the part thumbnail blogs never mention, and it is the useful part. The authors took a standard bottom-up saliency model — one that predicts fixations from colour, orientation and intensity contrast alone — and added a semantic channel that detects faces and text. The combined model predicted human fixations significantly better than the low-level model by itself. In other words, contrast is not sufficient to explain where people look. A quiet face beats a loud gradient.

What the study found	What it is often stretched into	What follows for a thumbnail
Faces drew 16.6× the fixations of matched control regions.	"Thumbnails with faces get more clicks."	A face reliably wins the first fixation. Whether it wins the click depends on what the face is doing and whether the video needs a person in it at all.
Text drew 11.1×, nearly as much as faces.	"Always add text."	Text competes with your face for the same reflex. One face plus five words is two magnets fighting; one face plus two words is a hierarchy.
The effect survived even when looking was penalised.	"You can force attention."	You can force the first fixation. You cannot force the second, and the second is where the decision lives.
Adding face and text detection beat low-level saliency alone.	"Contrast is everything."	Maximum contrast on a meaningless shape loses to moderate contrast on a meaningful object. Semantics outrank spectacle.

This is also the cleanest available explanation for why the two-element thumbnail keeps winning across niches. It is not minimalism as taste. A face and a short caption are the two strongest known attractors, and putting more than one of each into a 1280-by-720 frame makes them compete for the same involuntary first look.

Why "one glance" is not a figure of speech

Creators talk about having a split second. The experimental literature is more precise and more alarming. In work by Simon Thorpe and colleagues, photographs were flashed for 20 milliseconds and participants had to report whether an animal was present. Accuracy was 94%, with a median reaction time of 445 milliseconds. Follow-up work found that three weeks of training on a subset of the images did not make the processing any faster — novel scenes were categorised about as quickly as familiar ones. The speed is not a learned skill. It is the architecture.

The oft-quoted 150-millisecond figure comes from the same line of research but means something narrower: that is when the brain's electrical response to target and non-target images begins to diverge. It is the moment the category has been decided, not the moment a decision is acted on.

For a thumbnail the implication is uncomfortable but clarifying. Whatever your image is about at the category level — a person, a crash, a kitchen, a chart, a screenshot — has already been established before the viewer has read a single word or examined any detail you laboured over. The expensive parts of your design are evaluated second, if at all, and only because the gist bought them a hearing.

Which gives you a test rather than a slogan. Flash your own thumbnail at yourself for a fraction of a second — scroll past it fast, or shrink it to a quarter size and look away and back — and ask what category it registered as. If the answer is "a busy image" or "some text", the gist layer has failed and nothing downstream will rescue it.

The centre bias, and what it does and does not license

When people are shown an image, their first fixations cluster near the centre. The phenomenon has its own literature; Benjamin Tatler's 2007 Journal of Vision paper, "The central fixation bias in scene viewing", established that the bias cannot be explained away by motor habits or by where the image's features happen to sit — it holds even when the viewer's starting fixation is displaced. One surviving explanation is functional: the centre is simply the best place to stand to extract the gist of the whole scene, because it minimises how far anything is from the fovea.

Thumbnail advice turns this into "put the subject in the middle", which is close enough to be misleading. Two corrections matter.

The first is that the relevant centre is the centre of your tile, not of the screen. In a feed, every thumbnail gets its own small rectangle and its own first landing zone. The second is that centre-weighting applies to the discriminating element, not to every element. If the one thing that separates your video from the five next to it is a facial expression, that expression should be near the middle of the frame and large enough to resolve. If it is a number, the number goes there. Pushing the decisive detail into a corner, where the duration badge and the rounded crop eat into it, costs you the first fixation for no gain. We go through the geometry of this properly in the thumbnail composition guide.

Clutter has a definition, and it is measurable

"Simplify it" is the most-repeated thumbnail note in existence and the least actionable, because nobody says what simple means. Vision science does. Ruth Rosenholtz and colleagues at MIT built a measure called Feature Congestion, published in Journal of Vision, on a neat premise: a display is cluttered to the degree that it would be hard to add a new item that reliably captures attention. Their version 1.0 computes it from local variability in three channels — colour, luminance contrast and orientation — and it correlates with how well people perform visual search on complex images, most strongly when the thing being searched for is absent.

Read that as a specification. A thumbnail is cluttered when its colour varies a lot from patch to patch, when its light-dark structure varies a lot from patch to patch, and when its edges point in many different directions. Each of those is something you can fix deliberately:

• Colour variability. Hold the frame to two or three hues and let saturation do the work. A background photograph with eight hues in it is already congested before you add anything. Our guide to colour in thumbnails covers which pairs survive the feed.
• Luminance variability. Build one light-dark hierarchy and stick to it: subject bright against ground dark, or the reverse, consistently across the frame rather than alternating in patches. Local luminance chaos is what makes a thumbnail unreadable at a fifth of its size even when every individual element is high-contrast.
• Orientation variability. This is the one nobody manages. Radiating lines, tilted text, arrows at three different angles, a diagonal glow and a rotated badge are five orientation signals in one frame. Align them, or drop all but one.

The same measure explains why adding one more element to a thumbnail can reduce the visibility of everything already in it. You are not just adding an item, you are raising local variance across the whole frame, which makes the next thing you add harder to notice. Beyond a point, additions are subtractions.

Being looked at is not being chosen

The most useful corrective in this literature comes from a study that was not about thumbnails at all. An eye-tracking study of YouTube ad banners found that nearly every participant fixated on a banner at least once during a clip — and fewer than one in ten could correctly recall what the ad had said. Attention was captured almost perfectly. Nothing was retained.

That gap is the thing to hold on to when reading any attention claim. A thumbnail has to win four separate contests in order: it must be noticed, it must be understood, it must be chosen over the alternatives on screen, and the video behind it must then be worth the choice. Eye-tracking research speaks to the first and informs the second. It says nothing about the third and fourth.

Which is also why the one measurement that actually adjudicates your design is not a heat map but a test. YouTube's own A/B testing documentation is explicit that when you run up to three titles, thumbnails or title-and-thumbnail combinations, the option shown to everyone at the end is the one with the highest watch time share — not the highest click-through rate. The platform's reasoning is that packaging exists to set accurate expectations, and a version that wins clicks while losing viewers is not a win. A thumbnail optimised purely for attention is exactly the thumbnail that loses that test.

Where lab research stops being a guide to YouTube

Being specific about the limits is how you avoid the cargo-cult version of this material. Five gaps are worth naming.

The lab condition	The real condition	What changes
One image, alone, on a calibrated screen.	Twenty thumbnails competing in a scrolling grid.	Absolute salience matters less than relative salience. Your thumbnail competes against its neighbours, which is why whatever everybody in your niche is doing is a reason to do something else.
Fixed viewing distance, usually under a metre.	Phone at 30cm, laptop at 60cm, television at three metres.	Everything that depends on resolving detail degrades with distance. Living-room viewing sets the hardest constraint and is now a large share of watch time.
Naive viewers, no history with the stimulus.	Subscribers who recognise your face, your palette, your format.	Recognition short-circuits the attention contest entirely, which is the real argument for visual consistency.
A neutral task: look, or search for an object.	A mood: bored, winding down, looking for something specific, avoiding work.	Intent changes what counts as relevant. The same image wins in search and loses in browse.
Every viewer sees the same image.	Personalised surfaces, and now personalised packaging.	At Made on YouTube in September 2026, YouTube described dynamic thumbnails choosing between options for different audience segments. There may be no single winning image to find.

That last row is a genuine change in the problem rather than a caveat about methodology. If the platform can serve different viewers different packaging, the job shifts from finding the one best thumbnail to supplying a range that covers several plausible audiences without diluting what the video is. We covered the announcement and what it implies in our write-up of Made on YouTube 2026.

What about the saliency tools you can buy?

A number of products will take your thumbnail and return a predicted attention map. The honest description of these: they are models trained to predict human fixation data, and the better ones are genuinely good at it, because predicting fixations on static images is a well-studied problem with public benchmarks. What they return is a prediction of where the eye would go first.

That makes them a useful clutter-and-hierarchy check. If the predicted map lights up on your background texture rather than your subject, you have a real problem that is worth fixing. It does not make them click predictors, and any tool claiming to forecast click-through rate from an image alone is selling you a number it cannot know — CTR depends on the competing thumbnails, the traffic source, the title, the audience and the day, as our guide to what counts as a good CTR sets out. Use the map to find design faults. Use a watch-time test to pick a winner.

Translating the findings into decisions

Here is the whole of the above, reduced to the decisions you actually make while looking at an image.

Finding	Mechanism	Decision
Faces and text are near-reflexive gaze magnets.	Semantic channels outperform low-level contrast at predicting fixations.	One face, one short caption, maximum. Never two faces and a headline competing at the same visual weight.
Text draws nearly as strongly as a face.	Both recruit the same involuntary orienting.	Decide which one carries the idea, then subordinate the other in size and contrast rather than balancing them.
Gist is extracted in tens of milliseconds.	Rapid categorisation precedes detailed inspection.	Make the category unmistakable at a glance. Detail is a reward for the second look, never a requirement for the first.
First fixations cluster centrally.	Central position maximises gist extraction from the whole frame.	Put the discriminating element — expression, number, object — near the middle of the tile, clear of the duration badge.
Clutter is local variance in colour, luminance and orientation.	Feature congestion predicts search difficulty.	Cap hues at two or three, keep one light-dark hierarchy, align your edges. Audit by channel rather than by element count.
Fixation does not imply recall or action.	Attention capture and memory are separable.	Treat attention as the entry fee. Test for watch time, which is what the platform measures.

A sixty-second check before you publish

1. The flash test. Scroll past your own thumbnail at speed, or glance and look away. Name the category it registered as. If you cannot, the gist layer failed.
2. The small test. View it at feed size, not editor size. Anything that stops resolving is decoration you paid for.
3. The neighbour test. Put it beside five real thumbnails from your niche. Relative salience is the only kind that exists in a feed.
4. The greyscale test. Desaturate it. If the hierarchy collapses, you were relying on hue to do a job luminance has to do — which is also the fastest way to catch the pairs that fail for colour-blind viewers.
5. The orientation count. Count the directions your edges point. Over three, start aligning or deleting.

Then test, because attention is not the verdict

Everything above improves the odds that your thumbnail is noticed and understood. None of it tells you whether it was chosen. For that there is exactly one instrument that measures the outcome YouTube itself cares about: Studio's A/B test, which runs up to three options, decides by watch time share rather than clicks, and resolves within about two weeks. Results come back as a clear winner, a verdict that the options performed about the same, or no clear result — and when there is no clear winner, the option you uploaded first becomes the default, so upload your best guess first.

The scale of this is easy to miss. At Made on YouTube in September 2026, YouTube said creators had run more than 40 million A/B experiments on titles and thumbnails since 2024. That is a large body of evidence being generated continuously, almost none of it published — which is precisely why your own test results on your own audience are worth more than any general finding in this post. Our complete thumbnail testing guide covers how to design variants that can actually teach you something, and how much traffic a real answer costs.

The research in this post does one job for you: it narrows what to test. Instead of generating three arbitrary variants, you can vary one mechanism at a time — the face against no face, two words against five, a centred subject against an offset one, three hues against seven — and know that each contrast corresponds to something the visual system demonstrably treats differently. That is the difference between testing and guessing in triplicate.

It also sets a realistic ceiling on what design can do. Attention research explains why some images get looked at. It cannot make a video worth watching, and the watch-time criterion in YouTube's own test exists specifically to stop you optimising for the glance at the expense of the viewing. The mechanisms here are the floor of competence, not the ceiling of performance.

If the production side is what stops you — because designing three genuinely different variants per upload is the real obstacle for almost everybody — Thumblore exists to collapse that cost: describe the frame you want and generate variants that differ along one axis, so the thing you test is a hypothesis rather than whatever you had time to make. For the craft underneath, the psychology of clickbait thumbnails covers curiosity and expectation, and font choice for thumbnails covers the text half of the attention problem in detail.