Eight retention curve shapes and their causes
A diagnostic taxonomy for retention graphs: eight shapes, the look-alikes that get confused, and the discrimination test that tells them apart before you edit.
The reason retention advice rarely helps is that it is given about a shape the reader has not identified. "Tighten your pacing" is correct for one curve and actively wrong for three others. Before any edit is worth making, the graph needs a name.
A four-shape version is in audience retention analysis. This one covers the cases where two shapes look identical and have opposite repairs.
Read the axes honestly first
Three ground rules before naming anything, because most misdiagnosis happens at this stage rather than at the taxonomy stage.
Every curve declines. You are not judging decline, you are judging shape against your own channel median. MrBeast's leaked internal production document treats losing 21 million of 60 million viewers inside the first 60 seconds — around 35% — as an above-average result. Early drop-off is a rate to beat, not a defect to eliminate.
The opening is a different regime. Misclicks, wrong-audience impressions and pure feed churn all land in the first seconds and none of them are editorial problems. Judge the first 15 to 30 seconds separately from everything after it.
Compare like with like. YouTube's relative retention view — how your video holds against others of similar length — is closer to the comparison the recommendation system makes than any absolute percentage. On TikTok, length-banded floors circulate as rough working numbers: above 50% average watch under 30 seconds, above 40% for 30 to 60, above 30% past a minute. Those are practitioner benchmarks from analysed datasets, not platform-published thresholds.
The eight shapes
| Shape | Signature | Cause | Repair |
|---|---|---|---|
| Cliff | Steep drop inside the first 15–30s, then flat | Hook fails, or packaging mismatch | Cut the opening two seconds; open on a loaded frame |
| Early plateau at a low level | Drops fast, then holds steady | Title or thumbnail promised something else | Repackage. Do not re-edit |
| Slow bleed | Gentle, constant decline, no single event | Low information density, no re-hooks | Raise density; open loops on purpose |
| Mid-video cliff | Sharp drop between the 40% and 70% marks | One segment: a tangent, an ad read, a format break | Cut that segment |
| Step-down | Small drops at regular intervals | Weak transitions between chapters | Bridge each seam with a question |
| Valley | Dip, then partial recovery | A missing re-engagement beat that arrived late | Move the reveal earlier |
| Rewatch spike | Local bump above the surrounding line | Something replayable | Mine it — this is your next Short |
| Padding cliff | Drop at a length threshold rather than at a moment | A shorter idea stretched to clear a runtime target | Publish at true length |
Two of these are not problems. A rewatch spike is a signal to harvest, and a cliff of roughly your channel's usual size in the first 20 seconds is the tax everyone pays.
The three pairs that get confused
Each pair looks alike at thumbnail size and pulls in opposite directions.
Cliff versus early plateau at a low level. Both drop hard and early. The discriminator is what happens after the drop.
A cliff that keeps sliding means the video failed to give a reason to stay: a hook problem. A cliff that flattens into a stable line means a specific group left en masse and everyone else settled in. That is not a hook failure, it is an audience mismatch: the packaging pulled in people the video was never for. The first is an edit. The second is a title and thumbnail. Re-editing a healthy plateau usually removes the thing the remaining audience stayed for.
Read click-through rate alongside. High CTR with a savage early drop points at overpromising. Low CTR with the same drop points at a weak opening.
Slow bleed versus step-down. Both produce a curve that loses roughly the same total by the end. Zoom in.
A slow bleed has no events. Sample any 30-second window and the gradient is the same as any other window. The cause is density: not enough happening per unit time, no open questions carrying the viewer forward. The repair is structural, applied everywhere.
A step-down has events, they are just small. Drops of two or three points at recurring positions — usually the seams between sections. The cause is transitions that announce "that section is over" and give the viewer a clean exit. The repair is local and cheap: end each section on an unresolved question rather than a summary, so the seam stops being a door.
If you cannot tell which you have, count the visible inflections. Zero is a bleed. Four evenly spaced ones is a step-down.
Mid-video cliff versus padding cliff. Both are sharp drops in the back half.
A mid-video cliff is tied to content: something specific happens at 6:12 and people leave. Move the segment and the drop moves with it.
A padding cliff is tied to length: it shows up wherever the material ran out, and it recurs across videos at a similar position. One video with a drop at 5:40 is a segment problem. Five videos with drops between 5:15 and 6:00 is a programming problem — you are shipping longer than your ideas are.
The discrimination test is the back catalogue, not the single video. Pull your last eight uploads, mark each drop as a percentage of runtime and as an absolute timestamp. If the drops cluster on the timestamp axis, it is padding. If they scatter, each is its own segment.
The order to run the diagnosis in
Do it in this sequence, because early answers change what later steps mean.
- Check CTR before touching the curve. It sorts cliff from early plateau and stops you re-editing a packaging problem.
- Isolate the first 30 seconds. Name that regime on its own. The step-by-step for this is in diagnosing a weak opening, which joins the timestamp to what was actually on screen.
- Look for events in the body. Any drop steeper than the surrounding gradient gets a timestamp written down. No events means you are in bleed territory.
- Compare against three recent uploads. This separates a one-off segment fault from a recurring structural one, and almost nobody does it.
- Mark the spikes. Bumps above the line are pre-validated material. Repurpose by retention spike rather than by chapter — a rewatched 20 seconds is a hook already tested on a real audience, which beats anything you would have guessed.
Only then edit. If the steps agree, you have one edit to make. If they point at different faults, fix the earliest first: an opening fault contaminates every measurement downstream of it.
What the shape does not tell you
Two limits worth holding onto.
First, a drop after a completed payoff is not necessarily a failure. YouTube weights viewer satisfaction, not duration alone: leaving because you got the answer is different from leaving because it never came. The curve cannot distinguish those; watch the ten seconds before the drop.
Second, structure is a design input and not only a diagnostic output. The MrBeast document assigns explicit jobs by segment — minute one confirms the thumbnail's promise, minutes one to three carry visible escalation, and later blocks have their own briefs — with re-engagement beats scheduled deliberately rather than discovered afterwards. Reading curves tells you what went wrong last time. Assigning jobs to blocks is how you stop generating the same shape.
Turning a named shape into an edit
Naming matters more than it used to because the repairs got cheap. A mid-video cliff once meant a reshoot; on an EDL-based timeline it means removing a range from the edit list and re-rendering, with a free 480p preview pass to check the join before you commit — that pass carries a short per-user cooldown, and the final export is charged once no matter how many clips the timeline holds. A single bad beat inside an otherwise good segment is handled by a segment retake rather than a rebuild.
Step-downs are the exception that stays a writing problem: no editor fixes a transition that ends on a summary. That one goes back into the script.
Pacing repairs sit between the two. If the shape is a bleed, cut rate is usually part of it, and the per-platform cut rate numbers give you a target to edit against. The finding that matters there is variance rather than rate: constant cut frequency produces habituation, so a baseline punctuated by occasional bursts beats a metronome.
FAQ
How many videos do I need before these shapes mean anything?
Roughly eight to ten uploads of the same format. Below that you cannot separate a shape from one video's distribution luck, and the fourth diagnostic step has nothing to run on. Track it in a standing analytics review, not an occasional deep dive.
Can one video show more than one shape?
Routinely. A normal cliff in the first 20 seconds plus a step-down through the body is the most common combination. Fix in order of position: the earliest fault changes the population that experiences everything after it, so repairing a later fault first gives you a measurement you cannot trust.
Does this taxonomy work for short-form?
The event-based shapes do, compressed. Cliffs, valleys and rewatch spikes all appear inside 30 seconds. Padding cliffs mostly do not, since short-form has no runtime threshold to pad toward, and step-downs are rare with no chapter seams to weaken. On short-form, swipe-away rate carries most of the diagnostic weight anyway.
What about a bump at the very start of the curve?
That is a rewatch spike on the opening — viewers replaying the first seconds. Usually it means the hook was dense enough to need a second pass, which is a good sign. Check it is not confusion: if the same clip also shows a cliff immediately after, the replays were people trying to work out what they were looking at.