A retention graph is the closest thing short-form video has to a diagnostic. It shows, second by second, how many people were still watching. Read carelessly, it flatters you: the line starts high and the average looks respectable. Read properly, it tells you where you lost people and roughly why, which is the only information that changes the next edit.

The trap is that most creators read one number, the average, and then guess a cause. The average is a summary of a shape, and the shape is where the decision lives.

Start with the two numbers, not the curve

Before you look at the line, write down two figures: average view duration in seconds, and the percentage of the video that represents. Then note the split between people who kept watching and people who swiped away almost immediately. YouTube reports this directly for Shorts as viewed versus swiped away, and the other platforms expose an equivalent once a video has enough views.

Those numbers disagree constantly, and the disagreement is the useful part. A sixty-second video with a forty-second average is failing, because a third of its runtime is a place people leave. A twenty-second video with a fourteen-second average is doing better than it looks, because there was less room to hold anyone. Compare percentage viewed only against videos of similar length, never against the whole catalogue, where a three-minute explainer will always look generous.

The four shapes you will actually see

Almost every short resolves into one of four curves, and each implies a different fix.

The cliff. A near-vertical drop in the first two or three seconds, then a relatively flat line. The people who stayed were already committed; the people who left never entered. This is a hook problem, not a middle problem, and editing the middle of a cliff-shaped video is wasted work.

The slow bleed. A steady, gentle decline across the whole runtime. Nothing is broken and nothing is holding. This usually means the video delivers information at a constant rate with no peaks: no turn, no surprise, no specific reason to stay for the next ten seconds.

The mid-dip. A window where the line falls faster than the trend on either side, then recovers. Something concrete happened there. Find the timestamp before theorising about it.

The rewatch bump. A rise, sometimes a distinct second hump near the end, which means people went back. Very short clips under about fifteen seconds produce this almost automatically, so a bump there is not proof of brilliance. On a longer clip it is a real signal, and it is worth knowing which three seconds people replay.

A spike means rewatch, and rewatch is not always good

A rise in the curve means viewers returned, so the same seconds were watched more than once. That is usually read as a compliment. Sometimes it is a complaint: people replay because they did not understand a sentence, because a caption went by too fast, or because a number appeared on screen that needed reading twice.

The way to tell the difference is to watch the three seconds that produced the bump with the sound off, then again at normal speed, and ask what a stranger would be confused by. If the answer is nothing, you have a moment worth repeating. If the answer is that the text was unreadable, you have a comprehension problem wearing an engagement costume, and the fix is typography, not structure. Caption readability covers that fix in detail.

A dip repeated across videos is a habit, not an accident

One dip is an accident. The same dip at the same relative position in three videos is a habit, and habits are cheap to fix because you only have to diagnose them once.

Note where each of your last five videos loses the most, expressed as a fraction of its length. If every loss clusters around twenty percent of the runtime, look at what you habitually put there. Common answers: the title card, the first "so anyway", the transition where the music resets, or the point where you stop talking and start explaining.

The most common single cause at small scale is a cold open that hands off into a second introduction. The first five seconds earn attention; the next five spend it on setup the viewer already agreed to by staying. Cut the handoff and the curve usually improves without any other change.

Compare shapes, not averages

Two videos can share an average and have opposite problems. One holds everyone until a hard drop at the end; the other loses half the audience immediately and then holds the rest. Both average out somewhere in the middle.

So when you compare two videos, overlay the curves and compare them at matched positions: first three seconds against first three seconds, halfway against halfway. If you have no tooling for that, a rough method works. Screenshot each curve, note the percentage at zero, ten, twenty-five, fifty and seventy-five percent of runtime, and put those five numbers in a row. Patterns jump out of a row of numbers faster than out of a single average.

What the graph cannot tell you

The graph tells you where attention was lost. It cannot tell you why, and it never will. Asking the graph for a cause produces confident nonsense: "they left because the pacing dropped", when in fact the audio was clipping, or a caption sat under the platform interface and was never visible.

Two checks do most of the work of explaining a dip. First, watch the video on the platform it was published to, on a phone, with the interface visible, and see whether anything you put on screen is covered by it. That is so common that vertical safe zones exists to prevent it. Second, read the comments on that specific video for confusion, and read the transcript at the dip for a sentence that assumes knowledge the audience does not have. Only after those two checks is a structural theory worth forming.

The three timestamps worth writing down every time

Keep a log with one line per video and three timestamps: the point of first drop, the point of biggest drop, and the point people rewatched. Five minutes per video. After ten or fifteen videos you have a diagnosis of your own habits rather than a folder of curves nobody opens again.

The log also stops you from repeating an experiment you already ran. Most creators test the same hook variation four times without noticing, because the evidence of the first test lived in a screenshot they never looked at twice.

Turning one reading into one edit

The rule that keeps this useful: one reading produces one change. If a video lost people at second two and again at second eighteen, fix the hook in the next video and leave the middle alone, so the next curve is readable. Change both and you will not know which one worked, which is the same discipline that makes hook testing worth doing at all.

Retention is one of only a few numbers worth taking seriously. Which of the others earn their place, and which are decoration, is covered in which metrics are vanity and which are signal.

How to Read a Retention Graph Without Fooling Yourself