When a strong hook doesn't hold
A hook that wins the swipe and loses the viewer eight seconds later is a coherence failure, not a hook failure. The two-checkpoint read that separates them.
The most expensive misdiagnosis in short-form is treating a coherence failure as a hook failure. The clip stopped the scroll. People stayed for three seconds, four, five. Then somewhere around eight seconds the line falls off a shelf. The creator reads the overall retention number, decides the hook was weak, and rewrites the one part of the video that was working.
The distinguishing evidence exists and it is cheap to read. It just needs two checkpoints instead of one.
Checkpoint A and checkpoint B
Every video makes two separate sales. The first is the decision not to swipe, which happens in the first frames and is almost entirely a function of the opening image, the first spoken clause, and any text on screen. The second is the decision to keep watching once the viewer understands what the video actually is, which lands a few seconds later.
Checkpoint A is the swipe decision. YouTube Studio exposes this directly for Shorts in the Shorts feed tab as Viewed vs. Swiped Away. There is no official published pass mark, but the practitioner band that circulates is roughly 70% and up for a clip that is expanding its distribution, somewhere near half for a clip that is holding flat, and under 30% for a clip that is dead on arrival. Paddy Galloway's widely cited analysis of Shorts performance puts the healthy range at 70 to 90%, with collapse below 60%. Treat all of these as priors against your own channel median, not as thresholds handed down by the platform.
On long-form, checkpoint A is click-through rate. On TikTok and Reels there is no single exposed equivalent, so the proxy is the shape of the first bar in the retention chart relative to your own recent uploads.
Checkpoint B is the stay decision. For short-form, read retention at roughly eight to fifteen seconds. For long-form, read the 30 to 60 second mark. What you want is not an absolute number but the gradient between A and B.
The four combinations
| Checkpoint A (swipe/CTR) | Checkpoint B (8–15s / 30–60s) | Diagnosis | What to change |
|---|---|---|---|
| Low | Low | Hook failure | The first frame and first clause. Nothing else. |
| Low | High | Packaging undersells | Keep the video, rewrite the open and the title. |
| High | Low | Coherence failure | The body, or the hook's claim. Not the hook's craft. |
| High | High | Working | Mine it for a format, not a one-off. |
The third row is the one that gets misread. High A means the hook did its entire job: it earned attention. Low B means the video that followed did not honour whatever the hook implied. Rewriting the hook here makes the problem worse, because a better hook raises A and widens the gap.
TikTok for Business has said that 63% of its highest click-through videos establish a hook within three seconds. That number is often quoted as an argument for stronger hooks. Read carefully, it is an argument about timing, not intensity. A hook that lands in three seconds sets an expectation in three seconds. Everything after that is either payment or default.
The debt a hook takes on
A hook is a promissory note. It is useful to write down, in one sentence, exactly what the note says before you look at anything else. In practice there are five kinds:
- A question. "Why does every restaurant menu use this font?" The debt is an answer, and the viewer will start estimating how far away it is within seconds.
- A claim. "This costs eleven dollars to make." The debt is proof, itemised.
- A visual promise. A frame showing something mid-collapse. The debt is the collapse, on camera, not described.
- A scale promise. "I built the whole thing in a weekend." The debt is evidence of the scale, not a montage that could be anything.
- A tone or format promise. A cold open in a particular register. The debt is that register, sustained.
Now find the timestamp where the note is paid. Not gestured at, paid. In most coherence failures that timestamp is either past 30 seconds or does not exist, and the drop at checkpoint B sits roughly where the viewer finished estimating the distance and decided it was too far.
MrBeast's leaked internal production document assigns minute one a single job: confirm the promise the thumbnail made. That is the same mechanic at long-form scale. The opening is not there to be impressive; it is there to settle a debt taken on before the viewer pressed play.
Two repairs that point in opposite directions
If it is a hook failure (A low), the fix is upstream of the edit. Rebuild the first frame. OpusClip's analysis of 2,847 clips in early 2026 found visual-demonstration hooks to be the most-used subcategory of visual hooks, and visual hooks drawing more average views on TikTok than on YouTube over a seven-day window. The general lesson holds even if you discount the specific figures: open on a loaded frame, not on a person announcing what the video will be about. The single most common repair in retention work is deleting the first two seconds outright. Build a bench of openings up front rather than improvising one per upload; a hook pack generated before the video exists makes this a selection problem instead of a writing problem, and the first-three-seconds breakdown covers the families worth stocking.
If it is a coherence failure (A high, B low), you have two moves and they are both about the body:
- Move the payoff earlier. Find the first moment that pays the note and cut everything in front of it that is not required to understand it. This is usually setup the creator believed was necessary and the viewer did not.
- Downgrade the claim to what the video delivers. If the note says "eleven dollars" and the video proves "cheap", change the note. A slightly weaker hook with a body that pays it beats a strong hook that defaults, because the second one trains the audience to discount you.
The order matters. Try moving the payoff first, because it costs an edit. Rewriting the claim costs a re-record or a re-generation.
Running the repair without re-rendering everything
Both moves are timeline operations, which is why this loop is cheap on an EDL-based editor. The timeline is a set of instructions rather than a baked file, so pulling the payoff forward by twelve seconds is an edit to the edit list, not a new export of the whole video. You can watch the result through a free 480p preview pass before committing anything — it carries a short per-user cooldown, so it is for checking a specific change rather than scrubbing endlessly — and the final export is charged once regardless of how many clips the timeline holds. If the repair turns out to need one bad segment replaced rather than moved, segment retake handles that without touching the rest.
Then confirm it. Do not confirm on the same video: the numbers you already have are contaminated by the version you already published. Run the next pair the way any honest creative test runs — same posting slot on consecutive days, one variable changed, the metric named before you look, and at least three pairs before you believe the direction. The creative A/B protocol has the full stop-rule discipline, and the thing it protects against is exactly what happens here: peeking at a coherence repair after four hours and concluding the hook was fine all along.
One caveat worth holding. YouTube has been public about weighing viewer satisfaction, not just watch time, and a viewer who leaves because they got the answer is not the same event as a viewer who leaves because the answer never came. If your checkpoint B drop sits immediately after a clean, complete payoff, you may be looking at a video that worked and ended early rather than one that failed. Curve shape around the drop tells you which. A drop into a flat line after a payoff is a finished video. A drop into a continued slide is a default.
FAQ
How do I read checkpoint A if the platform doesn't expose a swipe metric?
Use the first visible bar of the retention chart against your own last ten uploads rather than a published benchmark. You are looking for whether this video started lower or higher than your norm, which is enough to place it in the table. On long-form, click-through rate is the cleaner stand-in because it is measured before playback rather than during it.
My checkpoint B drop is at 20 seconds, not 8. Same diagnosis?
Usually yes, but check what sits at 20 seconds first. Late drops often land on a specific event — a logo sting, a channel intro, a request for a follow before anything of value has been delivered. A pre-value ask is its own failure mode and it is fixed by moving the ask past the first payoff, not by editing the hook.
Is a coherence failure ever worth keeping?
Only if you are deliberately trading retention for reach on a single upload and you know that is the trade. It compounds badly. An audience that has been defaulted on twice starts swiping at checkpoint A on your next video, which converts a body problem into a hook problem for real.
Should I test the hook and the body at the same time?
No. The whole value of the two-checkpoint read is attribution, and changing both variables destroys it. Fix the body, run your pairs, then decide whether the hook still needs work. Testing hook categories against each other is a separate exercise with its own framework.