The Hook Spends Three Seconds In Three Layers

TikTok for Business put a number on it and the number has not moved since 2026-09-02: 63 per cent of top-CTR videos hook inside 3 s, and the average mobile decision takes about 1.7 s. That is the whole negotiation. The video does not die at 300 views because the content was thin. It dies because nobody stayed past the hook, and 300 views is what a dead intro looks like on a dashboard. 2026-09-02 is the day the hook library was written down as a spec instead of carried as taste, and the first line of the spec says the budget is 3 seconds and it is spent in 3 layers, in order, every time. Visual, then verbal, then context. Not 2 of the 3, and not in a different order because the edit felt better that way. The layers are ordered because the senses arrive in that order: movement and contrast register before a single word is parsed. 3 seconds is about 3 per cent of a 90 second cut and roughly 2 per cent of a 2 minute one, and it decides whether the other 97 per cent is ever seen. Retention above 70 per cent on the intro is the bar. Below 40 per cent stops distribution. 70 minus 40 is the entire distance between a video that gets pushed and one that gets buried, and both of them are sitting inside the same 3 seconds.

0.0 to 0.5 s is layer 1 and it is the layer most productions skip, because it is the one that has to work with the sound off. Movement, contrast, a bold text overlay, a face with energy. It must register before audio, and it must register on a muted phone held at arm's length. Visual hooks run roughly 100 times stronger than spoken ones for the simple reason that a viewer scrolling on autopilot is reading frames, not listening. The format is 3 to 5 words in a big bold font, centre or upper third, with 1 clear subject in frame and nothing else competing for the eye. The overlay has to sit out of the comment-box overlap and it has to be smaller than instinct suggests, because a frame with mashed text reads as work and work gets swiped. Text never covers the face or the subject. Contrast is the mechanism: a scroll-stopping visual paired with a contradictory text line stops the thumb harder than either half alone, and if the real frame is cluttered, the frame gets regenerated rather than rescued in the edit. There is 1 more rule and it is a motion rule. There has to be enough movement to hold attention and not enough to overwhelm, which in practice is a subject moving inside a stable frame, not a camera shaking and not a 12 shot burst inside half a second. Too much motion reads as panic and too little reads as a still image, and a still image does not stop a scroll. The colour of the text is pulled from the composition, never assigned from a palette. Red reads cheap and alarming even when the product is neither.

0.5 to 1.5 s is layer 2 and it gets exactly 1 sentence. A promise, a curiosity gap, or a problem. No opening greeting, no throat-clearing, no preamble about what is coming later. The sentence is staccato on purpose, because clipped sentences force clarity and pack value per word, and the video can expand to medium and long sentences only after the hook has done its work. The hook gets said aloud in the voice track inside the first 1.5 s, so the caption and the audio arrive together rather than 1 of them arriving late and the other carrying the load. Context comes first in this layer too, which is the single most common mistake in the format. The first frame text and the first spoken words carry the context, because a viewer who cannot place the subject in under a second has no reason to give the next 2. The strongest single framework here is the one that challenges a common belief: 45 degrees and carbs every day beats a list of things I avoid, and it beats it by a wide margin, because the 1st opens a question and the 2nd answers one nobody asked. Then the question test. What question does the line plant in the head of the 1 person this was made for? If the answer is nothing, or the question is hollow, the line gets rewritten rather than re-shot. A hollow question is the cheapest failure in the whole stack because it looks like a hook and behaves like a caption.

1.5 to 3.0 s is layer 3, and this is the layer that decides which viewer stays. Context lock means 2 things arriving in 1 line: who this is for, and what they walk away with. Topic clarity is the mechanism, and it works in both directions. The right viewer self-selects in, which is the point, and the wrong viewer self-selects out, which is the gift. A hook that keeps everybody keeps nobody, because the algorithm reads completion and the viewer reads relevance, and both of them are answering the same question in the same 1.5 s. The 2nd mechanism is speed to value. Short form is priced on a timer of about 4 s, and the value has to land inside the hook or immediately after it, never buried at the 76 per cent mark where the good part usually sits in a long edit. Frontload it. Value loop: context, then immediate value, then the next loop. Lead with the benefit or the pain rather than the topic, because wanting better sleep and needing magnesium is a hook, and magnesium being 1 of 21 core minerals is a spec sheet. When the subject is unfamiliar, wrap it in something already known so nobody has to feel behind before they can feel interested. And then the 1 line that separates a hook from an advertisement: obscure the reveal. Blur the key element, do not name the payoff, because the moment a viewer solves the question in their own head they have no reason to watch the answer. 8 checks grade all of this, and a clean safe zone plus an open loop is what a 7 plus looks like.

63 per cent of top videos hook inside 3 s, and the average mobile decision takes about 1.7 s.

The hook sits on top of an edit and the edit has its own numbers. 1 cut every 0.5 to 2 s, with most of them near 1 s. Beat-sync puts the cut on the music at 110 to 120 BPM, which is roughly 1 cut per 0.5 s, and the frames follow the grid rather than the other way round. Captions are essentially mandatory: burned in, keyword higher in the weight hierarchy than the rest of the line, 36 to 60 pt, and vertically placed where a thumb is not covering them. Velocity speed ramps are the 1 technique that does the most work in the format, with an S-curve ease on a snap zoom, and they are a property of the edit rather than the camera. The velocity lives in the cut, not in speeding up the voice track, which is why a video edited fast can carry completely unhurried narration and still feel urgent. Vertical 9:16 is the default and it is roughly 60 per cent of social video, so a frame composed for a 16:9 master loses its edges on the way to the only place it will be watched. Length is its own decision: 7 to 15 s outperforms longer for pure reach, and a short cut watched 2 times beats a 60 second cut watched once on every metric the platform publishes. Trend audio peaks in about 7 to 10 days and format trends run 3 to 4 weeks, so the useful window to jump in is the first 3 to 5 days after something starts moving. Arrive on day 12 and the format is already a costume. All of this is downstream of the hook, and none of it repairs one. A perfectly cut video with a dead first second is a well-built door with no handle.

7 frameworks cover the openings and they are all reusable: shock to logic, specific result, open loop, hot take, POV pain, speed-run tutorial, proof first. 8 audit checks grade them: context first, challenge a common belief, the question test, safe zone and declutter, first-impression mismatch, storytelling for saturated niches, do not let them solve it themselves, and visual plus text contrast. The published bar is 0 to 3 for broken, 4 to 6 for 1 element working, and 7 plus for all 3 elements with a clean safe zone and an open loop. Anything in the bottom band gets rewritten rather than optimised, because a broken hook has nothing to optimise. Under all of it sits 1 structure that has produced the largest results: the 3 step curiosity loop. Line 1 is the context lean, topic clarity plus lean in, with a common ground, a benefit, a metaphor or a plain surprising fact. Line 2 is the scroll-stop, and the strongest tool in it is a single contrasting word: but, however, yet, although. Line 3 is the contrarian snapback, which goes the opposite way from the lean while staying on topic. The screen is the least impressive part of the build, and the audio is the part that blows a mind, is what that shape looks like when it works. The probability that a hook written from taste alone clears all 3 elements on the first take is not zero. It is low, and the bill for finding out lands as a 300 view ceiling with no error message attached to it. The rule underneath the loop is 1 sentence long: open a loop, do not close it. Every closed loop opens the next 1. That is a storytelling commandment before it is an editing technique, and it is the reason a 2026 short at 3 seconds can hold a viewer harder than a 20 minute episode with 1 question and no engineering. The same discipline runs outside the feed. Write the bar down, 393 pages of it, built across twelve countries with 200 plus interns a year, and the work stops depending on the taste of whoever is holding the phone that day.