How to Read the YouTube Audience Retention Graph: Five Curve Shapes and What Each One Is Telling You

Most creators check retention the way you check a temperature. One number, once, after publishing: 47%, which is either fine or not fine depending on what the last video did. Then they close the tab. That habit throws away the single most diagnostic thing YouTube gives you, because the value of the retention report is not the average — it is the shape of the line and the timestamps where it bends.

Every other report in Studio tells you that something is wrong. Views are down, click-through rate is soft, impressions dried up. The retention graph tells you when. It is the difference between "this video underperformed" and "38% of the people who clicked were gone by 0:19, and the ones who stayed watched to the end" — two sentences that lead to completely different fixes. The first sends you to your editor. The second sends you back to the thumbnail.

This piece works through the report as a diagnostic instrument: what the curve is actually measuring, what the four labelled moments mean, the handful of curve shapes that recur across every niche and what each one is telling you, how to use the segment filters most creators never open, what a defensible benchmark looks like when the platform refuses to publish one, and the five questions the graph cannot answer no matter how long you stare at it.

What the line is actually plotting

Absolute audience retention plots, for every moment in the runtime, the percentage of viewers still watching at that moment. It opens at 100% — everybody who started the video was, by definition, watching at second zero — and falls from there. Read a point on it as a sentence: "60% at 3:00" means that of every hundred people who started, sixty were still there three minutes in.

Relative audience retention is the second view, and it answers a different question: how your curve compares with other YouTube videos of similar length. It converts an ambiguous number into a judgement. A 40% figure at the four-minute mark of a twenty-minute video is meaningless in isolation; relative retention tells you whether that is above or below what videos of that length normally do at that timestamp. Use absolute retention to find the problem and relative retention to decide whether it is a problem at all.

The graph lives at the video level only. There is no channel-wide retention curve, because averaging curves across different runtimes would produce a shape that describes nothing. You reach it through Studio's content list, the individual video, then Analytics — the retention report sits on the Overview and Engagement tabs. Retention data also lags: it typically settles a day or two after publishing, so the shape you read in the first hours of a launch is provisional and the early sample skews heavily towards subscribers and notification traffic.

The four moments YouTube labels for you

Under the curve, Studio marks specific moments in a report YouTube calls key moments for audience retention. Its Help Centre documentation defines four kinds:

Label	What YouTube counts	What to do with it
Intro	The share of your audience still watching after the first 30 seconds	Treat as the packaging score. Compare it across videos before comparing anything else.
Top moments	Stretches where almost nobody dropped off	Identify the format, not the topic. Whatever you were doing there, do more of it.
Spikes	Moments rewatched or shared, where viewership increased	Candidate Shorts, candidate thumbnails, candidate cold opens.
Dips	Moments skipped, or where viewers stopped watching entirely	Watch the ten seconds before the dip, not the dip itself.

The labels do not appear on everything. YouTube's documentation sets a floor — the video needs to be over a minute long and needs enough views before the report has anything statistically worth marking, commonly cited as around a hundred. On a small channel that floor is the reason the labels look inconsistent: the curve is always there, the annotations are not.

One clarification on Dips, because it is the label people misread most. YouTube's own wording covers two different behaviours: viewers who skipped a segment and viewers who stopped. A skip leaves the curve intact further on; an exit does not. Before you cut a section because it dipped, check whether the line recovers after it. If it does, the section was boring but survivable. If it does not, that section is where your video ends for a chunk of your audience.

Five curve shapes and what each one means

Across niches and formats, the same handful of shapes recur. Learning to name them turns a vague chart into a short list of suspects.

1. The cliff — a steep fall inside the first 10 to 30 seconds

Every video drops early; the question is how far and how fast. When the fall is severe and then the remaining line flattens and holds, the diagnosis is nearly always a mismatch between what the packaging promised and what the opening delivered. The people who left were never the audience for this video — they clicked something the video was not. The ones who stayed were well matched, which is why the rest of the curve looks healthy.

This is a packaging problem wearing an editing problem's clothes, and it is the single most common misdiagnosis in the report. If your intro figure is weak across many videos while the mid-video curve is fine, the fix is upstream of the timeline: the thumbnail, the title, and the first sentence that either honours them or does not. Our guide to the first thirty seconds covers the craft of that opening in depth.

2. The slide — a smooth, gentle decline with no sharp features

This is the normal shape, and it is the one that needs the least intervention. A steady decay means nothing in particular is broken; people are peeling off at the rate people peel off. The lever here is not repair, it is length. A slide that reaches the end at 45% on an eight-minute video is a strong result; the same gradient on a twenty-five-minute video will land somewhere far lower, and the honest question becomes whether the last ten minutes were earning their place. That trade-off is the subject of our piece on how long a YouTube video should be.

3. The step — a sudden vertical drop mid-video

A cliff at 6:40 with a flat line either side is the easiest fix in the report, because it points at one specific thing. Go and watch from 6:20. The usual culprits are structural rather than qualitative: a sponsor read that arrives before the video has earned it, a topic change with no bridge, an unannounced shift in format, a long tangent, a section that repeats what was already said. Steps rarely come from bad content. They come from content that the viewer could not see the point of continuing through.

4. The sawtooth — repeated small dips and partial recoveries

A jagged curve where the line drops and partially recovers is the signature of skipping, not leaving. Viewers are scrubbing past segments and rejoining. This is common on tutorials, recipe videos and anything with chapters, and it is not necessarily a failure — it can mean your video is a reference that people navigate rather than a story they follow. The productive response is usually to accept the behaviour and serve it better: clearer chapters, tighter segments, the answer earlier. Fighting the skip by hiding the information deeper just converts skippers into leavers.

5. The rise — the line climbs

Somebody scrubbed back. Rises cluster around demonstrations, reveals, a number on screen, a before-and-after, a line that landed. The instinct is to feel good about them and move on. The better move is to treat every rise as raw material: the frame at a rewatch peak is, empirically, a moment your audience found worth seeing twice, which makes it an unusually strong candidate for a thumbnail or a Short. This is one of the few places where analytics hands you a creative decision rather than a verdict.

The first thirty seconds is a thumbnail report

Here is the reframe that makes the retention graph worth opening weekly. Views are the product of three separate things: how often YouTube showed the video, how many people clicked, and how long they stayed. Click-through rate measures the middle step. The intro section of the retention curve measures whether that click was honest.

A thumbnail can fail in two directions and the metrics tell them apart cleanly. A thumbnail nobody clicks shows up as low CTR and a normal retention curve. A thumbnail that overpromises shows up as healthy CTR and a first-thirty-seconds cliff. The second failure is the expensive one, because it spends impressions on viewers who bounce, and bounced viewers are a worse signal to the recommendation system than viewers who never clicked at all. Read those two numbers together or you will keep optimising the wrong half of the funnel — our guide to what counts as a good CTR makes the same argument from the other side.

There is a legitimate version of the curiosity gap and a dishonest one, and the retention curve is where the difference becomes measurable rather than a matter of taste. A thumbnail that creates a question the first thirty seconds answers will hold its viewers. A thumbnail that creates a question the video never answers will not, and the graph will show you exactly how many people worked that out and how quickly. We covered the mechanics of that gap in the psychology of clickbait thumbnails; the retention report is the audit trail.

The segment filters almost nobody opens

Under the retention chart, Studio lets you split the curve by viewer group. This is where the report stops being descriptive and starts being diagnostic, and it is routinely ignored.

• Subscribers versus non-subscribers. If subscribers hold and non-subscribers cliff, the video is fine and the packaging is over-promising to strangers. If both cliff, the opening itself is the problem.
• New versus returning viewers. A curve that only works for returning viewers means the video assumes context that new arrivals do not have — usually an in-joke, a running format, or a recap you left out.
• Organic versus paid. If you run ads, comparing the two curves separates audience quality problems from content problems.

Traffic source is the other cut worth making, and Studio's Advanced Mode is where you combine it with retention. A video that holds search viewers and loses browse viewers is behaving exactly as you would expect: search traffic arrives with a specific question and stays until it is answered, while browse traffic arrives on a whim and leaves on one. Judging a single average retention figure across a mix of both sources tells you very little, for the same reason a blended click-through rate does — a point we made at length in YouTube traffic sources explained.

Retention, AVD and average percentage viewed

Three numbers describe the same underlying behaviour and get used interchangeably, which causes more confusion than it should.

Metric	What it is	Best used for
Average view duration	The average number of minutes watched, among people who played the video	Comparing videos of different lengths on the same channel
Average percentage viewed	That same watch time expressed as a share of the runtime	Comparing videos of similar length
The retention curve	The percentage still watching at every moment	Finding the timestamp where something went wrong

The relationship is simple arithmetic: a ten-minute video with 50% average percentage viewed has an average view duration of five minutes. What matters is that the two summary figures move in opposite directions when you change length. Cut a video from twenty minutes to twelve without removing anything people were watching and your percentage viewed goes up while your total watch time may go down. Neither number is lying; they are answering different questions, and only the curve tells you which minutes were actually the problem.

What counts as good, honestly

YouTube does not publish retention benchmarks. That absence is not an oversight — a benchmark that ignored niche, format, length and traffic mix would be actively misleading. Into that gap have poured a great many vendor numbers, and the striking thing when you line them up is how much they disagree. One widely circulated set of figures calls 65–75% strong for videos under five minutes, 50–60% for five to ten, 40–50% for ten to fifteen and 35–45% beyond that. Another vendor's guide puts "good" at roughly 30% average percentage viewed, or 60% still watching past the first thirty seconds. Both are presented with equal confidence.

They can all be right, because they are averaging different populations. Treat any of these as a rough sanity check and nothing more:

Runtime	Third-party "strong" range for average percentage viewed
Under 5 minutes	65–75%
5–10 minutes	50–60%
10–15 minutes	40–50%
15 minutes and up	35–45%

These come from analytics tools and creator surveys, not from YouTube. The two comparisons that actually hold up are the ones the platform gives you for free: relative retention, which grades you against videos of similar length, and your own last ten uploads, which grades you against the only audience that matters. If your intro figure moved from 62% to 71% across five videos, that is worth more than any published range.

How retention feeds the recommendation system

Retention is not a dial the algorithm reads directly, and the folk model of "high retention equals promotion" oversimplifies what YouTube has said about its own system. In a conversation with YouTube liaison René Ritchie, the platform's senior director of growth and discovery, Todd Beaupré, described a recommendation system that predicts what an individual viewer wants next and weights watch time by how satisfied viewers appear to be — measured through survey responses, whether they come back, and what they do after your video ends.

The practical translation is that retention is evidence rather than currency. Nobody is rewarding you for a number; the system is trying to predict whether the next viewer will be glad they clicked, and your curve is one of the better predictors available. That is also why a short video people finish can travel further than a long one they abandon, and why an artificially stretched runtime tends to backfire twice — once in the curve and again in the satisfaction signals behind it.

It shows up most concretely in thumbnail testing. Studio's Test & Compare runs up to three thumbnails on one video and decides the winner on watch-time share rather than raw clicks, over a test window of up to two weeks. That design has a consequence worth sitting with: a thumbnail can win the click and still lose the test, because the viewers it attracted left early. Retention is the tiebreaker built into the feature. Our guide to running a Test & Compare covers the mechanics, including how long to leave a test alone before reading it.

Shorts do not have this graph

On Shorts, the equivalent question is answered by a different pair of metrics: how many people the Short was shown to in the feed, and how many chose to view rather than swipe away. The key moments report is built for videos over a minute; for a thirty-second Short the meaningful window is the first two or three seconds, where the decision is binary and instant.

Vendor benchmarks for the viewed-versus-swiped ratio circulate widely and should be treated with even more caution than long-form ones, since the population of Shorts is enormous and wildly heterogeneous. What is safe to say is directional: the average percentage viewed on Shorts runs far higher than any long-form benchmark, simply because the runtime is short enough to finish by accident, and a swipe-away rate that sits well above your own channel's norm is the clearest signal that the opening frame is not stopping the scroll.

Two changes that affect how you read the curve in 2026

First, view counting. From 24 August 2026, YouTube counts a public view from the very first frame of playback across long-form, Shorts and live, with no minimum watch threshold. The stricter, older definition survives inside Analytics as engaged views. This does not change viewer behaviour or your retention curve's shape, but it does change the denominator on anything you calculate yourself from public view counts, and it means comparisons that straddle that date need care.

Second, Studio itself is moving. YouTube has been rolling out a redesigned analytics experience in which the Analytics tab is renamed Insights, charts open into a more detailed Advanced Mode, and an assistant called Ask Studio answers plain-language questions about your channel by pulling from multiple reports at once. Along with title A/B testing and the reworking of the Trends tab into a research destination, the practical effect is that the retention report is being surfaced in more places and summarised for you more often. A summary is not a diagnosis; the curve still repays reading yourself.

A twenty-minute retention audit

A routine beats an occasional deep dive. This one fits between uploads:

1. Open your last five videos and write down only the intro figure — the share still watching after thirty seconds. Five numbers, one column.
2. If those five numbers are all weak, stop. The problem is packaging or openings, and nothing further down the timeline is worth analysing yet.
3. If they vary, take the best and worst and compare the first minute of each, second by second, against the script. Name the difference in one sentence.
4. On the worst performer, split the curve by subscribers and non-subscribers. If only non-subscribers cliff, rewrite the thumbnail and title, not the video.
5. Find the largest mid-video step in each video. Note the timestamp and what happens twenty seconds before it. Look for a pattern across all five — most channels have one recurring structural habit that costs them the same drop every time.
6. Find the biggest rise. Export that frame as a thumbnail candidate and that segment as a Short.
7. Write one sentence about what you will do differently in the next video. One. A retention audit that produces six action items produces none.

Five things the graph cannot tell you

Honest limits matter as much as the reading technique, because most bad decisions made from this report come from asking it a question it was never able to answer.

• Why anyone left. It records departures, never motives. Every "because" you attach to a dip is your hypothesis, and it is worth testing rather than believing.
• Whether a low-retention video was a failure. A video with a weak curve and enormous reach can produce more watch time, more subscribers and more revenue than a tightly held small one.
• Anything reliable in the first hours. The early sample is dominated by subscribers and notification traffic, and the data itself takes a day or two to settle.
• Whether the drop was your fault. Phones die, feeds interrupt, people watch a recipe halfway and start cooking. A share of every curve is life, not craft.
• What to make next. Retention grades what exists. It has no opinion about the video you have not made, and channels that optimise it too hard converge on the same safe, high-retention format and stop growing.

The habit worth building

Read the intro figure on every video, the shape once a week, and the segment splits whenever something surprises you. That is the whole practice. The creators who improve fastest are not the ones with the most sophisticated analytics stack; they are the ones who have converted "this video didn't do well" into "we lose 30% of clicks in the first fifteen seconds, and we have lost them there four times running" — because the second sentence has a fix attached to it.

More often than not, that fix lives before the timeline rather than inside it. When the curve holds after the first half-minute but the first half-minute keeps collapsing, you are not looking at an editing problem. You are looking at a promise your packaging made that the video did not keep, and the cheapest place to fix it is the image. If that is where your curves keep pointing, Thumblore generates thumbnail options from your video's own frames in seconds, which makes the loop of test, read the intro figure, and try another angle short enough to actually run each week.

Then close the loop properly: use the intro figure, not just click-through rate, to judge which thumbnail won, and let Studio's own test decide it on watch-time share. If your curves keep collapsing at the same point, start with the first thirty seconds; if they collapse everywhere at once, the wider diagnosis in why your views dropped is the better place to begin.