Looking at the Camera Underperforms: Presenter Data From 2,499 TikTok Videos

Three pieces of advice this data does not support

We analysed 2,499 TikTok videos with a video model that records how the person on camera behaves — where they look, what their face is doing, whether the delivery reads as scripted or spontaneous, how many people are in frame.

Three near-universal pieces of creator advice came out backwards:

• Look at the camera. Videos with no eye contact outperformed direct eye contact on every metric.
• Smile. Smiling was among the worst-performing expressions. Serious was among the best.
• Just be authentic. "Genuine" was the most common classification and the worst performer.

Disclosure. FYPNow sells the analytics product that produced this data. We have an

obvious interest in you finding it credible, which is why the method, the sample bias and

the reasons these findings might be artefacts are all stated below rather than buried. The

figures are published under CC BY 4.0 — check them.

Before the tables, the boundary on what you can conclude, because it matters more here than in most studies: these are model classifications, not ground truth. A video model deciding whether a delivery is "genuine" or "mixed" is making a judgement call, and there is a section at the bottom showing you a place where that taxonomy visibly leaks.

Eye contact

Direct eye contact — the thing every guide tells you to do — is last on both metrics, about 21% below "none" on engagement.

The likely reason, and it is not "eye contact is bad". A video with no eye contact is frequently not a talking-head video at all. It is a screen recording, a voiceover over B-roll, a montage, a text-on-screen explainer. Those are different formats with different jobs, and some of them are intrinsically more saveable — you save a screen recording of a process, you rarely save someone's face.

So the honest reading is not "stop looking at the camera". It is that talking to camera is not the baseline format it is treated as, and the formats that do not require it are not handicapped. If you have been avoiding TikTok because you do not want to perform to a lens, 820 videos here say you do not have to.

Facial expression

Smiling and excited — the two most obviously positive expressions — are the two worst performers. Serious and animated are near the top.

There is a coherent mechanism available. A serious delivery signals this is information, which is the setup for a save. A smiling delivery signals this is pleasant, which is the setup for a scroll. Enthusiasm is cheap and viewers appear to price it accordingly.

"Just be authentic" is the worst-performing mode

The most common classification by a factor of four is also the worst by a wide margin — 0.45% against 0.86%, a gap of about 90%.

Be careful with this one. The most plausible explanation is not that sincerity repels viewers. It is that "scripted" and "mixed" are proxies for production effort: a video the model reads as scripted probably had a written hook, planned cuts and a deliberate structure. "Genuine" may largely mean "unplanned", and unplanned content underperforming is not a surprising result at all.

Read that way, the finding is much less spicy and much more useful: the authenticity everyone recommends is not a substitute for planning it. The best-performing bucket is "mixed" — a planned video delivered naturally — which is what good creators have always done.

Solo beats a double act

Two people on camera performed worst on saves — below having nobody there at all. Solo is both the dominant format (83%) and the best performing.

The result that mattered least

Save rate differs by 0.02 percentage points and engagement goes the other way. This is a null result, and we are reporting it because null results are the ones studies quietly drop.

Perceived confidence was the single least predictive attribute we measured. If you are holding off publishing until you feel confident on camera, nothing here suggests that wait is buying you anything.

Where the taxonomy visibly leaks

Look again at the expression table. The best-performing category on both metrics is "other" — 1.11% save rate, 9.23% engagement, comfortably ahead of every named category.

That is a tell, and it is worth naming rather than hiding. When the catch-all bucket outperforms every labelled option, it usually means the labels are not capturing the thing that matters. Sixty-five videos got sorted into "other" because the model could not fit them into smiling, serious, animated, neutral, excited or none — and whatever those videos are doing, they are doing it better than anything the taxonomy has a word for.

We could have dropped that row and the post would read cleaner. It stays because it is the most honest indicator in the study of how much to trust the other rows: the classifications are useful, directional, and not measurements of a physical quantity.

What we would actually do with this

Stop treating talking-to-camera as the default. It is 37% of the sample and the worst-performing eye-contact mode. Screen recordings, voiceovers and text-led formats are not second best. Drop the performed enthusiasm. Smiling and excited both underperformed serious and animated by wide margins. Deliver like you are telling someone something useful, not like you are pleased to be there. Plan it, then deliver it naturally. "Mixed" beat both pure extremes. That is the practical version of the authenticity advice. Film alone. Solo beat two-up by 62% on save rate. Do not wait to feel confident. It was the least predictive thing we measured.

Method

2,499 TikTok videos carrying presenter analysis, each processed by the same multimodal model, which classifies eye contact, facial expression, perceived authenticity, confidence level and the number of people on camera directly from the footage. Save and engagement rates come from each post's own metrics; engagement rate is (likes + comments + shares) / views.

Tables reporting rates cover the 2,492 of those videos that also carry a view count, since a rate needs a denominator. Categories with fewer than 50 videos are excluded. All figures are medians, because view counts are heavily skewed and a single outlier moves a mean anywhere.

The sample is videos FYPNow users chose to track, which skews toward content someone already thought worth watching. Every comparison here is correlational, several have confounds we have named, and the classifications are a model's judgement rather than ground truth.

Related

• 2,584 TikTok videos analysed: the most common hook isn't the best one — the opening seconds
• 89% of TikTok videos have no call to action — overlays, captions and CTAs
• The anatomy of a TikTok video — length, pacing and audio
• Find viral TikTok videos in 4 steps — run this on your own niche