Find the segment behind a mid-video cliff

A drop in the back half of a video feels like a pacing problem and almost never is. Pacing problems produce a gentle, eventless slide. A sharp drop between the 40% and 70% marks is an event: something specific happened, a large group of people responded to it in the same way within a few seconds of each other, and the thing that caused it is sitting on your timeline with a start and an end.

Global fixes do not repair local faults. Tightening the whole video by 20% because of a cliff at 5:40 shortens the good parts and leaves the bad segment in, slightly faster. The procedure below is the body-half counterpart to the weak-opening diagnosis, and it is a search, not a rewrite.

Locate the window, then watch before the drop

Retention graphs are bucketed, so the timestamp you read off the chart is a neighbourhood rather than a point. Convert it deliberately.

Take the percentage where the line breaks, multiply by runtime, and then widen it into a window. For a ten-minute video, one percent of runtime is six seconds. At 25 frames per second — the default on a Versely timeline, and worth knowing because it is 25 and not 24 — that is 150 frames per percentage point. A drop that reads as "around 57%" is therefore a window roughly 30 seconds wide, not a moment.

Write down three things before you open anything:

• The drop percentage and the runtime it converts to.
• The magnitude — how many points the line loses across the break.
• The gradient before and after. This decides the diagnosis in the next section.

The instinct is to scrub to the timestamp and watch forward. That shows you the moment the audience was already gone.

Leaving is a two-part action: a decision, then a delay while the viewer waits to see whether the thing they disliked is going to end. On long-form that delay commonly runs ten to twenty seconds; on short-form, a couple. So the cause sits before the break.

Practical version: start playback 30 seconds before the drop and watch to 10 seconds after it, once with sound and once muted. The muted pass is not optional — it catches the visual half of the fault, the half you went blind to after four hours in the edit.

Name the fault from a short list

Mid-video cliffs come from a small number of causes. Work down the list in order; the first match is usually the answer.

Fault	What you'll see in the window	Repair
Tangent	The video stops advancing its main thread and services a side point	Cut the tangent whole
Ad read or sponsor block	A format break with a different register	Move it behind a payoff, or shorten it hard
Pre-value ask	A subscribe or follow request before the viewer has been paid anything	Move it past the next payoff
Repetition	A point already made, restated because the script needed a bridge	Cut the second instance
Format break	Camera, energy, aspect or audio character changes for no narrative reason	Smooth the seam or cut across it
Question answered	The title's promise was fully paid just before the drop	Possibly nothing — see below
Energy dip	Delivery flattens; the narration stops carrying stakes	Retake the lines, not the segment

The last two need care.

Question answered is the one that looks like a fault and may not be. YouTube has talked publicly about viewer satisfaction as a ranking input, not duration alone, and a viewer who leaves because they got what they came for is a different event from one who leaves because it never arrived. The tell is the gradient after the break. A drop into a flat line means a cohort completed the video and the remainder is watching for something else. A drop that continues sliding means people left unsatisfied and took their neighbours with them.

Ad reads and CTAs are worth separating because placement, not existence, is usually the fault. An end-screen CTA is served to the smallest cohort the video will ever have, so the working window is immediately after a payoff, framed around the value just delivered. The same CTA before a payoff is a reliable cause of a drop.

Make exactly one cut

One edit per diagnosis. Two changes and you cannot attribute the result.

Three shapes of repair, in increasing cost:

1. Remove the range. Cut the tangent or the repetition out entirely and check that the join makes sense — usually it needs a sentence trimmed on either side rather than a hard splice.
2. Move the range. Ad reads, CTAs and asides mostly do not need deleting, they need relocating behind the nearest payoff.
3. Replace the range. If the fault is delivery rather than content, the segment is right and the take is wrong.

On an EDL-based timeline all three are operations on an instruction list rather than regenerating the whole video, which is what makes this loop practical weekly. Check the result through the free 480p preview pass before committing — it carries a short per-user cooldown, so treat it as a verification step rather than a scrub tool — and the final export is charged once regardless of clip count. For the replace case, a segment retake swaps a single bad stretch without disturbing anything around it.

Resist the fourth option, which is all of the above plus a general tightening pass. It feels productive and it destroys the experiment.

Confirm on the next episode, not this one

This is the step that gets skipped, and skipping it is why creators repeat the same fault for a year.

You cannot confirm the repair on the video you repaired. Its retention data blends everyone who watched the old version with everyone who watched the new one, weighted by whenever the swap happened, on top of a distribution curve that has already moved past its early spike. Whatever number you read will confirm whatever you already believe.

Confirm on the next episode of the same format instead:

1. Keep the format fixed. Same length band, same structure, same slot. You are testing one structural change, so everything else must hold.
2. Apply the same repair by design. If the fault was a tangent at 57%, the next script has no tangent there. If it was a pre-value ask, the ask moves behind the payoff.
3. Pre-commit the read. Name the checkpoint — usually retention at the equivalent percentage — and the direction that counts as a win, before publishing.
4. Look once, at the checkpoint. Peeking early is the dominant failure mode in every creative test, and early data skews toward whatever the feed did in the first hours.
5. Repeat across three episodes before believing it. A single episode's retention is dominated by which audience the recommendation system happened to sample.

If three consecutive episodes with the repair show a flatter curve at that position, the fault is fixed and the repair belongs in your script template permanently. If they alternate, the segment was not the cause and you are back at the fault list with the second-most-likely match. The general stop-rule discipline is the same one that governs any creative A/B test.

When it is not one segment

Three signals that you have a structural problem wearing a segment problem's clothes:

• The drop appears at a similar absolute timestamp across several uploads. Different content, same minute mark, is a length problem, not a content problem.
• The window contains nothing unusual. You have watched it four times, muted and unmuted, and there is no event. That is a bleed with a slightly steeper patch, and the repair is density, not excision.
• There are four or five small drops rather than one large one. That is a step-down at chapter seams, repaired by rewriting transitions to end on open questions rather than summaries.

Naming the shape before starting the search saves the search entirely in those cases; the full shape taxonomy is the faster front door. If the fault does turn out to be rhythm rather than content, the per-platform cut rate figures give you a concrete target to edit against instead of a vague instruction to speed up.

FAQ

How big does a drop have to be before it's worth investigating?

Compare gradients, not absolute points. If the line loses noticeably more across the break than across a comparable window elsewhere in the same video, it is an event. A two-point dip in a curve already losing two points per equivalent window is just the curve.

The window before my drop contains a sponsor read. Do I remove sponsors?

No, you move them. Placement is the variable that matters: behind a payoff, framed around the value just delivered, keeps the cost low. A sponsor read dropped into the middle of an unresolved thread asks the viewer to hold a question through an interruption, and a good share of them will not.

What if the segment I need to cut is the one the client asked for?

Show them the curve with the timestamp marked. A drop with a magnitude attached is an easier conversation than an opinion about pacing, and it usually converts into a shorter version of that segment placed later rather than an argument about whether it stays.

Does this procedure work on short-form?

Yes, with the clock compressed. The decision-to-exit delay shrinks to a couple of seconds, so the window before the drop is two to four seconds rather than thirty, and the fault list narrows — tangents and ad reads mostly vanish, leaving format breaks, repetition and the answered question.