Hooked,
or skipped.

Anyone who runs short-video content for a living has been here: ten videos, same account, same week, made with the same care. One breaks 200,000 views. The other nine stall at a few hundred. The tempting conclusion is that the feed is random — an algorithm coin flip.

This is not the thousand-and-first hook template article. Templates age fast and can't answer the question that's actually worth money: how much does the opening decide — enough to reorganize how you produce? We went looking for answers in our own operating data and in two academic studies that happened to measure the same thing from very different angles. What follows is what we've found so far. We are not a research institution — just a team that works with this data every day. People in this industry have certainly dug deeper; corrections are welcome.

The real shape of playback data

Creator dashboards speak the language of averages. Average watch time: 11 seconds. Average completion: 43%. These numbers quietly paint a picture of a mildly interested viewer drifting away somewhere in the middle.

Open the dashboards and look at the curves one by one. Videos matching that 'drifting away' picture are extremely rare. What appears instead is two extremes: a video either loses most of its audience in seconds or carries most of them through. Plot enough videos by watch depth and the distribution splits into two peaks with a valley between them — very few videos sit in the middle.

This is not a quirk of our account pool. In 2024, Snap and the Chinese University of Hong Kong measured the same shape on 90,000 short-feed videos — every one viewed by at least 2,000 real users, a scale that eliminates noise. Both their headline metrics (overall watch depth and five-second survival rate) showed the same bimodal pattern.

Think about what two peaks mean. If viewers were flipping a coin at every video, per-video averages would cluster in the middle — one peak, not two. Two peaks mean the opposite: thousands of strangers, faced with the same video, reach the same verdict with high consistency. A video is not playing a lottery. It is being voted on, and the vote rarely comes back close.

A video with '45% average completion' is rarely watched halfway by everyone. It's usually two audiences: one that left in seconds, and one that stayed to the end.

(To be precise: the papers measure the bimodal pattern across videos. That a single mid-scoring video is also internally split into quick leavers and finishers is our judgment from reading client retention curves — pull one of yours and you'll see it.) Either way, this changes what 'improving a video' means. When outcomes cluster at two poles, making the whole thing 10% better is not a meaningful goal. The goal is moving it from the left peak to the right one.

The weight of the first five seconds

Everyone says the first three seconds matter. The question worth answering is: how much, exactly? We ran the numbers on our own data.

The TikTok accounts we manage for brands produce second-by-second retention curves every day. We pulled 10,000 authorized videos — 10 to 60 seconds long, each with at least a thousand views, all from the official API — and measured the cliff:

More than half the audience is gone before the five-second mark. But the cliff itself isn't the main point — the next number is.

Controlling for video length, the rank correlation between five-second survival and total watch depth on our accounts is 0.83. In plain terms: just by looking at how the first five seconds perform, you can largely rank how deeply the entire video gets watched.

And one more number closes the loop: among the top third of our videos by opening strength, only 3.6% ended up in the bottom third of watch depth. 'Strong opening, mid-video collapse' barely exists. The loss happens at the opening.

You might object: of course they correlate — anyone watching deeply obviously survived second five. But the two numbers are not the same measurement. A video could hold everyone past the opening and then collapse at second ten; it would score high on survival and low on depth. A correlation this high says that almost never happens. The fate of a video forks at the opening. After that, the path is mostly set.

Nobody is ever moved by a beautiful ending they didn't stay to see.

There is a second, colder reason to care: the platform is watching the same numbers. This research originally addresses a recommendation-systems problem called cold start — when a new video arrives, the platform gives it a small test audience and quickly decides whether to spend more traffic on it. The paper's explicit goal is to predict engagement from content alone, before the audience arrives. So your opening is graded twice: once by human thumbs, and once by models trained to anticipate those thumbs. (Note: the paper studies Snapchat's cold start specifically. That major platforms use similar logic for new videos is my judgment from watching traffic allocation — not their data.)

(Data note: our numbers come from July 2026 snapshots, authorized accounts, official TikTok Business API. This is our account pool, not a platform-wide sample.)

What actually predicts watching

A common and expensive reaction when a video underperforms: upgrade the camera, refine the color grade, add a gimbal. The researchers ran a pilot study that directly tests this instinct. They took two professional video-quality models — UVQ and DOVER, both trained on human ratings of clarity, sharpness, and visual appeal — and asked: how well do quality scores predict real watch time? Videos were compared only against others of similar length, to keep it fair.

The correlations came back between 0.07 and 0.31. Knowing that a short video is beautifully shot tells you almost nothing about whether anyone keeps watching. 'Quality' as the industry has measured it — resolution, stability, visual appeal — barely registers in real viewing behavior. The reason is the bimodal chart again: most viewers leave before your production values take the stage.

So what does matter? Working on creative content every day, we had a rough sense — meaning over sound, sound over polish — but it was a hunch until this study measured it. The researchers built a model that reads the content itself, adding one signal at a time and tracking what each addition does to prediction accuracy:

The total gain: everything added together lifts prediction by just 0.07 over frames alone — and nearly half of that (+0.032) comes from one row: the machine learning to 'see' what's happening on screen. Each row maps to a concrete judgment about where effort should go:

1. Visual basics — 0.625, the baseline. Just the picture — quality, what's in frame, how it moves — already reaches 0.625. A large share of viewing behavior is predictable from frames alone.
2. Sound & music — up to 0.636. Sound adds a real but small step. It carries pace and mood — viewers feel the tone before they understand the content. Worth having; don't overrate it.
3. Title & description — up to 0.651. The creator's own text helps more than sound. The title sets what viewers expect coming in, and matched expectations keep people.
4. Auto-caption — up to 0.657. A one-sentence machine summary of the video: small gain. One line throws away too much. The machine's real value is not the sentence it writes (see the next kept row).
5. Speech transcripts — dropped (−0.004). Counterintuitive: transcribing everything said in the video made prediction slightly worse. The drop is tiny — a weak signal on its own. But paired with the fact that only ~30% of these videos had usable speech, and the keep-or-skip decision usually lands before anyone finishes a sentence: don't stake your opening on what gets said.
6. Machine 'seeing' — up to 0.689, largest jump (+0.032). The main course. The captioning model, while learning to describe videos, first builds an internal representation of the picture: who is in frame, what they're doing, what the scene is. The researchers skipped its text output and fed that internal representation directly into prediction — one step of +0.032, more than double the next-best signal (titles, +0.015). This is the data behind the article's central point: what decides watching is the meaning on screen, not the finish.
7. Aesthetics — up to 0.696, final garnish (+0.007). Composition and visual appeal help — as the smallest kept step. Craft earns points; it doesn't buy entry.
8. Emotion labels — dropped (−0.006). Tagging frames 'happy/sad' also made things slightly worse. At minimum: coarse single-frame emotion labels didn't help. For those who like 'emotional hooks' as a category — at this granularity, this data doesn't support it.

The fairest comparison: even after retraining the strongest quality model (DOVER) on the same real-viewing data, it still loses to the content model (DOVER retrained: 0.635; content model: 0.696). Content signals win not because the other side was mis-tuned — they carry more information. (Don't confuse this 0.70 with the earlier 0.926: that was behavior predicting behavior; this is content predicting behavior before anyone has watched — much harder, hence the lower number.)

One detail worth noting for anyone without a following: the model uses no account data at all. No follower count, no creator history — deliberately, because it's built for brand-new videos in cold start. The predictable part of engagement lives in the content itself.

What the eyes do in the first seconds

When we review content before publishing, one check comes first: does the opening frame give the eye a single place to land? That was editor's instinct until an eye-tracking study turned it into something measurable.

The study: about 20 viewers per ad watched 151 thirty-second TV ads with eye trackers recording exactly where each person looked, second by second. Two concepts are all you need to read the results. A saliency map is a heat map of where eyes land on a frame. Entropy is a single number summarizing scatter: low entropy means every viewer's gaze converged on the same spot; high entropy means twenty people looked in twenty different places.

The extremes are clear. Frames in the most-focused tenth almost always feature one thing: a centered face, a person, a product held to the camera. Frames in the most-scattered tenth have competing focal points — several objects, text in multiple corners. To a scanning eye, three points of interest add up to none.

The most practical finding concerns cuts. These ads averaged 15.6 scenes in 30 seconds — a cut roughly every two seconds. At every cut, measured gaze scatter jumps while viewers hunt for a new anchor, then settles once they find one. Every cut re-runs the audition. A fast-cut opening is not intrinsically a strong opening; it's a series of chances to lose people, each of which must be re-won.

Scenes rarely half-fail: a scene with high average scatter tends to stay scattered throughout. If a shot diffuses attention, the fix is not trimming a frame or two — it's recutting or replacing the scene.

The researchers also trained a model that predicts these attention maps from video alone — no eye tracker needed. Its predictions correlate 0.51 with measured scatter: not high, but enough to flag which seconds of an unreleased piece deserve re-examination. The direction matters more than the number — attention diagnosis is moving from post-mortem to pre-publish check.

Now put the two bodies of evidence side by side. One shows outcomes are binary: hooked or skipped, almost nothing in between. The other shows the opening's mechanism is binary too: gaze either locks onto an anchor, or scatters. Different data, different countries, different screens — but they describe the same first seconds from opposite ends, and they agree.

Six production rules

These are not hook templates — templates are everywhere and age fast. These are production-process-level rules, the kind that stay true after this month's formulas stop working. They come directly from the evidence above, and we use them in our own workflow. Take what's useful.

1. Grade your first five seconds as their own product. The opening is not the start of the video; it's the gate the rest must pass through. Review it in isolation, at feed speed, on a phone — muted first. Many viewers scroll with sound off, so sound is a bonus, not the pass bar (the feature ladder confirms this: sound adds only a small step). Acceptance test: muted, can a stranger say what this video is about and why they should care? In analytics, make 3-second and 5-second retention first-class numbers. Pull your last 20 videos, rank by 5-second retention, then by completion — watch how little the order changes.
2. Diagnose by curve, not by average. 'Average 45% completion' tells you nothing about your typical viewer — it's a blend of two audiences. Pull the retention curve. If the cliff is inside the first five seconds, you have a hook problem — this article is the prescription. If viewers survive the opening but drain mid-video, you have a different problem: pacing and story, not attention capture. That's rarer (recall 0.83), and worth its own treatment — we'll write about it next. The curve is in your creator dashboard. The average hides which disease you have; the curve shows it in one glance.
3. One anchor per opening frame. The eye-tracking data is clear: one centered subject — a face, a product, a single line of text. Negative example: product + price sticker + logo + lyric captions in the same frame = four focal points = zero. Positive example: one hand holding the product dead center, one line of eight words or fewer. Across formats — talking-head: face plus a one-line conclusion, not a channel intro; tutorial: finished result up front, not a greeting; story: the frame of peak conflict, not the first beat of setup. Design the first frame like a poster, not a collage. (The evidence here is from TV-ad eye tracking; take the mechanism, don't transplant the numbers.)
4. Every cut is a re-audition. Attention re-anchors at every scene change, and short-form content averages a cut every two seconds. Carry continuity across cuts — subject position, motion direction, a visual through-line — so the eye lands instantly. If a scene scatters attention, recut it whole: the data says scenes rarely half-fail.
5. Version the hook, not just the video. If the verdict lands in five seconds, that's where testing should happen. One body with three or four different openings is a better experiment than four different videos — cheaper, and it isolates the variable that moves outcomes. Where do alternatives come from? Your body usually contains a better opening than your opening: cut the two densest seconds from the middle and put them first; open on the finished result instead of the build-up; take the strongest single sentence and make it the first frame's text. Paid accounts can A/B openings as separate creatives. Organic accounts: space versions a day or two apart and re-edit the body slightly — never repost identically, or the platform may flag it as duplicate content. No budget? Show three openings to five people and ask which one would stop their thumb.
6. Spend the afternoon on hooks, not on color grading. The feature ladder is explicit: meaning and sound carry the prediction; aesthetics add the final one percent. Color grades, lens upgrades, and smoother gimbals are real craft — sitting behind the gate. Same afternoon, five alternative hooks versus one more round of color grading — the data has already answered.

To close the rules, here is our own data — it speaks to the 'testing' half of rule 5: winners exist, and modest test volume finds them.

Over the past 90 days, we ran a TikTok matrix campaign for one consumer-brand client: same product, same pool of accounts, 36 creative formats, 962 posts total, publishing deliberately balanced at 16–34 posts per format. (A 'format' here means a whole creative pattern — how the first image opens, the copy structure, the pacing. Unboxing, comparison, mini-story: three different formats.)

Method honesty: this is not a clean opening-only experiment — the formats differ as whole creative patterns. But in our campaigns, this lesson repeats week after week: winners exist, the gap between a winner and the field is not tens of percent but multiples, and twenty-odd posts per format is enough to find them. (Numbers from our backend, July 7, 2026, as recorded; client identity withheld.)

What these studies don't prove

Still — two independent teams, two very different instruments (a platform's logs, a lab's eye trackers), one consistent picture: the verdict happens early, and attention either locks on or leaks away. Our own dashboards show the same thing. Our sample is limited, but it has been enough to shape our production process around one assumption: if the verdict happens in the first seconds, that is where creative work — and creative testing — should concentrate.