Your hook is a frame, not a line
The opening frame carries more hook load than the opening line. What clip-level data shows, the sound-off case, and a first-frame audit for your last 30 posts.
Most hook advice is written as if the viewer is listening. It gives you sentences: "Stop scrolling if you...", "Nobody tells you that...", "Three things I wish I knew before...". Good sentences, and on a muted feed a large share of your audience never receives them.
The frame arrives first and it arrives unconditionally. There is no autoplay setting, no headphone state, no ambient noise level that stops a viewer from seeing pixel one. Everything else in your opening is conditional on something you do not control. That asymmetry is the whole argument, and it changes what you should be optimising: not the first line of the script, but the first still.
What the clip-level data actually shows
OpusClip published an analysis of 2,847 clips that used a visual hook, dated early 2026. Visual-demonstration openings — the clip starts on the thing happening — were the largest subcategory at n=1,316, ahead of fast-paced-scene and shock/surprise openers. The page does not publish view averages for those other two subcategories, so it does not support a performance ranking among them. What it does support is simpler: demonstration is the visual-hook pattern creators actually use most, and "interesting frame" is not the same instruction as "busy frame."
The same analysis reported visual hooks earning roughly 1.6x the average views on TikTok as on YouTube over a seven-day window — 632 against 394 average views. That is a platform-placement difference for one hook family, not evidence about hook families, and it comes from a vendor-analysed dataset rather than a platform disclosure. Treat it as a prior about where to spend visual openers first.
The one platform-sourced figure worth holding onto is TikTok for Business's own line that 63% of its highest click-through videos establish the hook inside three seconds. Three seconds is roughly one shot at the cut rate TikTok rewards — so in practice the hook window is one to two frames' worth of decisions, not a paragraph.
The sound-off case, with real numbers behind it
Two independent studies make the frame the load-bearing channel rather than a nice-to-have.
Extreme Reach's 2026 five-country survey (US, UK, France, Spain, Germany; more than 3,000 respondents) found 49% of US viewers always or often watch with captions on, a behaviour the study described as consistent across screens and genres. Coverage of the same study put always-or-often caption use on short-form at 39%. An earlier Verizon Media and Publicis survey of 5,616 US adults found that 80% of respondents said they were more likely to finish a video when captions were available, and that 50% said captions matter because they watch with the sound off.
Put those together and the picture is not "some people mute." A substantial, stable share of your audience is reading and looking, never listening. Your opening line reaches them only if it is also on screen — at which point it competes with your frame for the same two seconds, and the frame wins because it needs no parsing. This is why a sound-on strategy is a separate decision from a hook strategy: sound-on design is about what audio adds for the people who have it, and hook design has to work for the people who do not.
The first-frame audit
Here is the exercise. It takes about ninety minutes for thirty posts and it is the single highest-yield hour of analysis available to a short-form account.
- Pull the last 30 posts you published on one platform. One platform only — mixing them contaminates the comparison, because the feeds seed differently.
- Extract frame one of each. Turning a video into stills does this in bulk, which is the difference between a ninety-minute exercise and an afternoon of scrubbing timelines.
- Lay all thirty stills out in a grid at thumbnail size. Do not watch the videos. This is the point of the exercise: you are seeing what a scrolling viewer sees.
- Classify each frame into one bucket: demonstration in progress, person mid-sentence, product at rest, text card, environment/establishing, logo or brand plate, or dark/fade-in.
- Attach one number to each row — average watch time at a fixed 24-hour checkpoint, not lifetime views. Views are dominated by later distribution and will tell you about the algorithm's mood, not your opening.
- Sort by that number and look at the buckets. With thirty posts you will not get significance. You will get a ranking, and usually one bucket sitting conspicuously at the bottom.
The result most accounts get on the first pass: the dark/fade-in and logo buckets sit at the bottom, the demonstration bucket at the top, and there are more frames in the bottom two buckets than expected. That gap between what you thought you were shipping and what the grid shows is the actual finding.
Four openings that fail the grid test
These fail for the same reason: at thumbnail size, with no audio, they carry no information.
| Opening | What the viewer gets in frame one | Fix |
|---|---|---|
| Fade from black | Nothing. You have spent your most valuable frame on a transition. | Start on the loaded frame. Fades are for endings. |
| Logo or brand plate | A brand they have not agreed to care about yet. | Move it to the last two seconds, alongside the ask. |
| Establishing wide | Context for a story they have not been given a reason to enter. | Open tight on the anomaly, widen later. |
| Face mid-breath | A person about to say something. The promise is entirely in audio. | Open on the face reacting to something already visible. |
The face case surprises people, because talking-head content works. It works when frame one shows a state — an expression that implies a story — not the neutral instant before speech begins. The practical adjustment is to start the clip a second later than feels natural.
What to put in frame one instead
Three properties, in priority order.
An unresolved state. Something is mid-way through happening and its outcome is not visible. A hand halfway through a fold. Liquid in the air. A screen showing a number that is obviously wrong. Unresolved beats impressive, every time, because impressive resolves.
One subject, large. Thumbnail-size legibility is a hard constraint. If a viewer has to find the subject, they have already scrolled. This is why the demonstration bucket wins in practice: a demonstration naturally frames one thing doing one thing.
A claim in text, if the frame alone cannot carry it. Text overlay is a legitimate part of the frame, but it is a second channel, not a caption of the first. If your overlay transcribes your first sentence, you have spent two channels on one message. Keep it out of the bottom ~15% and top ~10% where platform UI sits, and give it its own idea. Adding a text overlay is a timeline operation, so you can try three versions of the claim over the same frame before committing.
When you generate the opening rather than shoot it, this stops being an edit problem and becomes a prompt problem. You can ask an AI video generator for the exact unresolved moment instead of hoping a take contains one, and the free 480p preview pass in the editor lets you check the first frame at thumbnail size before you spend a final render on it. It carries a short per-user cooldown, so batch the checks rather than tapping it repeatedly.
Make the frame a production input, not an edit decision
The order most accounts use is: make the video, then find a hook in it. That produces openings that are the least-bad frame available rather than the frame you wanted. Reverse it and the whole pipeline changes shape — you decide the opening still first, then build the video as its payoff. That is the argument behind generating a hook pack before the video exists, and it is the version of hook work that survives contact with a content calendar.
The line still matters. It converts the viewers the frame already stopped, and the first three seconds is where the two have to agree. But it is second in the sequence, and treating it as first is why so many well-written hooks underperform on muted feeds.
FAQ
Does this apply to YouTube Shorts as well as TikTok?
The sound-off argument applies everywhere, but the metric you audit against differs. Shorts exposes "Viewed vs. Swiped Away" in the Shorts Feed tab of Studio, which is a more direct read on the opening than average watch time is. Practitioners treat the healthy band as roughly 70–90%, with real trouble below 60% — a convention from creator-side analysis rather than a published YouTube threshold, so calibrate it against your own channel median before you act on it.
How many posts do I need before the audit means anything?
Thirty gives you a ranking, not a result. It is enough to spot a bucket that is clearly failing and to notice how much of your output falls into openings you would not defend. For an actual comparison between two opening types, you need a paired test rather than a retrospective audit, because your thirty posts differ in topic, timing and length as well as in frame one.
What about series content where viewers already know the format?
Returning viewers are a different population, and a recognisable opening frame genuinely helps them. The problem is that the same frame reads as nothing to the non-followers the feed is testing you on, and non-follower reach is where growth comes from. The workable compromise is a consistent element — a colour, a position, a recurring object — inside a frame whose main subject still changes every time.
Should the first frame match my grid thumbnail?
Matching them is fine and often good. What you should not do is design frame one for the grid at the cost of the feed: the feed is where the impressions are, and the grid is where people who already found you look.