Everyone told me the same thing when I started building thumbnails full time. Put a face on it. Big eyes, open mouth, high contrast, and the click-through rate takes care of itself. So we did that for a while, because it is easy advice to follow and it is not wrong often enough to get you fired.
Then we started actually testing it. Not on client channels first, because that is a rude place to run an experiment. We tested on our own channels, where the only person losing views was me. Over a stretch of months we shipped paired versions of the same video packaging: one with a human face doing the emotional work, one where the face was gone entirely and the idea carried the frame.
The short version: faces won more often than they lost, but the margin was much thinner than the advice implies, and the losses were not random. They clustered. Once we saw where they clustered, the way we design changed.
What we were actually testing
A lot of thumbnail "testing" is not testing. Somebody swaps an image, the video is three days old, the algorithm has already decided who it is showing it to, and the CTR number moves for reasons that have nothing to do with design. I have made that mistake and reported it as a result. Not proud of it.
So we kept it narrow. Same title. Same video. Same publish window. The only variable we allowed was whether a human face was the dominant element in the frame. Everything else, the type treatment, the color contrast, the crop tension, the number of objects, stayed as close to identical as we could get it.
That is harder than it sounds, because removing a face leaves a hole. You cannot just delete the person and call the second version a control. You have to fill that space with something that does comparable work. Usually that meant an object shot larger, a number, or a before and after split. The honest way to describe our test is not "face versus no face." It is "face carrying the emotion versus an idea carrying the emotion."
The result that surprised me
On our own channels, in the education and process categories, the face versions won on click-through most of the time. Fine. Expected. But the gap was often in the range of a few tenths of a percentage point, not the doubling that gets quoted in every packaging thread.
What I did not expect: on several videos the face version pulled a higher CTR and a lower average view duration. People clicked the face, arrived, and left faster. That is the pattern that made me stop treating CTR as the scoreboard.
A face gets you the click. It does not get you the watch. If the face promises a personality and the video delivers a tutorial, you have just paid for traffic that leaves.
This is the same trap I wrote about in more detail when a high CTR thumbnail actually costs you. The number goes up, the channel gets worse, and it takes weeks to notice because the dashboard is congratulating you the whole time.
Where faces lost, consistently
Three clusters. Every one of them makes sense in hindsight, which is how you know a result is probably real rather than noise.
Comparison and versus content. When the video is A against B, a face in the middle is a third character nobody asked about. The two things being compared want the whole frame. Every time we shrank the face to make room for both objects, we ended up with a small face doing nothing and two objects that were still too small. Removing the face entirely won.
Number-led and result-led videos. If the hook is a figure, the figure should be the biggest thing in the frame. A face next to a big number splits attention between two loud elements. In our tests the no-face version with a huge number and one supporting object beat the face version more often than not.
Channels where the host is not yet the draw. This is the uncomfortable one. A face works partly because viewers recognize it. Below a certain familiarity threshold, a stranger's face is just a stranger. We saw this on a newer channel of ours: the face versions underperformed until the channel had been publishing consistently for months, and then the same treatment started winning. Nothing about the design changed. The audience changed.
I want to be careful here. These are results we saw on specific channels in specific niches at specific sizes. They are not laws. A cooking channel and a finance channel do not behave the same way, and a finance channel around 80k subs does not behave like the same channel at 400k. Anyone selling you a universal face rule is selling you a rule that has not been tested against your channel.
The part of the face that matters
When face versions did win, it was rarely because the face was there. It was because of what the face was doing.
We kept scoring our own thumbnails on three things: eyeline, jaw and brow, and hand involvement.
Eyeline mattered most. A face looking at the camera and a face looking at the object in frame do very different jobs. Looking at the camera creates a direct address, which suits opinion and story content. Looking at the object directs the viewer's attention to the thing that matters, which suits process and review content. The versions that failed most often were the ones where the eyeline went nowhere in particular, because then the face is just decoration taking up 40 percent of a very small rectangle.
Jaw and brow beat mouth. The exaggerated open mouth reads as generic now. It has been used so heavily that it carries almost no information. A tight jaw, a raised brow, a slight lean, those read as a specific reaction to a specific thing, and specificity is what survives at 320 pixels wide on a phone.
Hands helped more than I expected. A hand holding, pointing, or covering something adds a second focal point that supports the face rather than competing with it. When we added hands to weaker face versions, several of them flipped from losing to winning.
What this changed in how we design
We stopped starting with the face. That is the actual change.
The old habit was to shoot a reaction, cut it out, and then build a thumbnail around it. The face becomes the given and everything else negotiates for the space that is left. The new habit is to decide what the single idea is, decide what the biggest object in the frame should be, and only then ask whether a face makes that idea land harder. Sometimes the answer is yes. Often the answer is that a face would be a fourth element in a frame that can hold three.
Because every thumbnail we make is custom per video and never a template, this decision gets made fresh each time. That is slower. It is also the entire point. A template applies the face rule to a video that might be one of the three cases above where the face rule loses.
This is the same reasoning behind how we think about YouTube packaging strategy generally: the thumbnail is not a standalone graphic, it is the second half of a sentence the title started. If the title already names the person, the thumbnail probably does not need to show them. If the title is a number, the thumbnail should not be arguing with it.
How to test this yourself without wasting a month
If you want to run this on your own channel, a few things I would insist on.
- Test on videos you expect to be average. Testing on your outlier is how you learn nothing, because the outlier was going to travel regardless.
- Change one thing. If the no-face version also has different type and a different background color, you have not tested faces, you have tested a whole new thumbnail.
- Watch retention alongside CTR. A face version that wins CTR and loses the first 30 seconds is not a win.
- Give it enough impressions. A few thousand is a rumor, not a result.
- Write down what you predicted before you look. It is the only defense against reading whatever you wanted to see.
Most channels do not need a formal A/B setup to learn this. Four or five paired attempts across a quarter will tell you more about your specific audience than any general advice will. We usually build a variant for this reason, which is a small add-on at fifteen dollars on top of a pack, but you can absolutely do it manually with a swap after 48 hours if you are careful about timing.
If you are running this across a team of editors and a channel manager, the harder problem is not the test, it is keeping the results somewhere they survive staff changes. I went into that separately in the piece on building a thumbnail workflow for a team, because a test result nobody wrote down is a test you will run again next year.
The cost side of being wrong
One more thing worth saying, because it is the reason we bother testing at all.
A thumbnail that underperforms does not just lose the clicks it lost. It teaches the algorithm something about who your video is for, and that lesson sticks past the moment you fix the image. I laid out the fuller version of this in the cost of a bad thumbnail, and the face question is a good example of it. Slapping a face on a versus video because that is the received wisdom is a decision that costs you real distribution on a video that would have done fine with two clean objects.
This is also why I am unconvinced by fully automated generation for this specific decision. Tools are getting good at producing a face on a background. They are not good at deciding whether this video wants a face at all, because that decision needs context about the channel, the audience's familiarity with the host, and what the last six videos looked like. I compared the two approaches honestly in AI thumbnail generators versus designers, including the places where the tools genuinely win.
If you want to see how these decisions shake out across different niches, our work is the fastest way to get a sense of it. And if the videos you are packaging are long-form interview or podcast episodes, the thumbnail question changes again, because the face in that format is usually two faces and the framing rules are different. Our sister studio handles podcast editing and clips if that is the format you are in.
What I believe now
Faces are a strong default, not a rule. They work best when the audience already knows the face, when the eyeline is doing a job, and when the frame is not already carrying a number or a comparison. They work worst when they are added out of habit to a frame that was already complete.
The useful skill is not knowing that faces raise CTR. It is knowing the four or five situations where they do not, and having the discipline to leave the face out on those videos even though it feels like you are skipping a step.
We are still testing. I expect some of this to change again in six months, and I would rather say that than pretend a few months of data settled it. If you want us to run the same process on your channel, contact is the fastest way to start.