A Pitch Sample's Retention Curve Drops. Recut the Video—or Change What You Measure?

A twelve-person audience makes an odd graph. Each session is worth a little over eight percentage points, which means the dip you are staring at is not a trend. It is a person, or two. That is the arithmetic of most pitch samples: a curve assembled from a short list of recipients, several of whom you can name and most of whom you could email before lunch.

This is why "should I recut?" is usually the wrong first question. A retention curve is good at one thing — telling you where something happened in a video. It is bad at the thing you actually need, which is why. A fall at 0:30 and a rise at 1:52 are coordinates. They are not a transcript of anyone's thoughts, and the shape alone cannot separate a viewer who gave up from a viewer who skipped ahead to the part they had been told about.

So the useful order is: fix the conditions the curve came from, point at the moment, hold at least two explanations for it, then choose the smallest thing that could tell them apart. Sometimes that thing is an edit. Often it is a question, or a note about who was sent what.

What the curve is a curve of

Before reading the shape, write down what produced it. Six items, and the short version is that if you cannot fill them in, you are interpreting a document you cannot date.

• The upload and the cut. Which version is this curve from? If two cuts of the sample exist, the report belongs to one of them, and it is easy to spend a week debating a graph that came from the wrong export.
• Duration and the observation window. A three-minute sample read over nine days is not the same object as the same sample read over nine hours. Early data from a link sent to fourteen people is shaped by when each person happened to be at their desk.
• The metric as the platform defines it. Relative retention, absolute retention and "views" are different quantities with different denominators, and they move differently when someone rewinds.
• Who received it, when, and what they were told. This is the item most often missing, and the one that most often explains the curve.
• What recipients were asked to do. "Watch the sample" and "have a look at the reveal around the middle" produce two different videos' worth of behavior from the same file.
• How the link travelled. Sent individually by email, forwarded inside a company, shown on a screen in a meeting, or pasted into a group chat — four different distributions.

The last three matter more than they look. A pitch recipient is not the eventual series audience. A recipient is doing a job: deciding whether to take a meeting, whether the show fits a slate, whether to ask a colleague. That job involves time pressure, prior context and internal politics that no viewer of episode three will ever share. A curve from fourteen recipients describes fourteen people doing diligence. It cannot carry a statement about audience demand for the series.

Two limits apply at this size. At twelve sessions, one person is worth a little over eight points, so a five-point "decline" is a fraction of a person — below the resolution of the instrument. And a report that flags no key moment is not a report in which nobody rewatched anything. It means the tool surfaced nothing. Absence of a flag is a weaker statement than absence of the behavior, and if that difference matters to your decision, it belongs in the follow-up, not in the conclusion.

Worth checking in the report you actually have: the provider documents viewing retention by segment type, which may let you break the curve apart. Whether the segments available to you separate "stopped" from "skipped ahead" is something you have to verify in the interface you are looking at, not something the help page promises.

To make this concrete, here is an invented sample. Coin Drop is a fictional unscripted series about a two-person workshop that restores jukeboxes. The fictional pitch sample is three minutes, sent as an unlisted link to fourteen named recipients at four companies over four days. Over the following nine days the report shows twelve viewing sessions, ten of them longer than thirty seconds, from nine distinct recipients. No analytics account, video or recipient was consulted to build this example; every number in it is made up to work through the method.

Point at the second, not at the adjective

The provider's own definitions are the place to start, because they are already more careful than most of the conversation around them. YouTube's help page on measuring key moments for audience retention covers a spike with watching, rewatching and sharing, and a dip with abandoning or skipping. It also allows that repeat viewing can accompany material people went back over because it was unclear, not only because they liked it. (Definitions as documented in September 2026; they are the provider's, not a finding about your sample.)

Note what that leaves open. Skipping and stopping are both dips. From the curve alone, the recipient who bailed at 0:40 and the recipient who jumped from 0:40 straight to the reveal cast nearly the same shadow. The provider's labels are honest about the ambiguity. The graph does not resolve it for you.

Here is the invented trace, with the moment each row corresponds to:

Relative retention in this table means the share of the starting sessions still viewing that second. The whole table is invented.

Three features, nothing more. A steep loss across the introduction. A plateau across the repair montage — the section a producer would usually call slow. A narrow bump of exactly two sessions at the reveal, gone inside twelve seconds.

The plateau deserves as much attention as the drop, and gets none. But be careful how far you take it. The six sessions still live at 1:06 are not a random sample of the original twelve. They are the ones who chose to stay, and they will keep choosing to stay, which means a flat stretch late in a video partly describes who is left rather than what the material is doing. The montage is not proven watchable. It is proven survivable by six people who had already survived forty-four seconds of talking.

The rule that keeps this honest: mark the moment, describe the behavior, stop. "Retention falls forty-two points between 0:00 and 0:44" is an observation. "Viewers found the introduction boring" is a claim about the inside of someone's head wearing measurement clothes. That second sentence is how recuts get ordered for the wrong reason.

Two stories that fit the same dip

Take the two features separately, and give each one a genuine rival.

The introduction drop. Story one is content: forty-four seconds of a host explaining a show is a long time to ask for trust before offering evidence, and the sample delays its own payoff. Story two is viewing task: the sample went out as a bare link, so the first minute absorbed every "open it, see what this is, come back later" session, plus any recipient who had been told there was a reveal and skipped forward to find it. Both stories predict a fall in the first minute. They predict different things about the fall's shape, about whether it would repeat, and about where those sessions went next.

The reveal spike. Story one is value: the machine playing is the thing people wanted, so they watched it twice. Story two is clarity: the moment was quick, or the host talked over it, and recipients went back to work out what had happened — the same repeat behavior with the opposite editorial meaning attached. Story three is route: a recipient forwarded the link with a start time, or someone opened the sample mid-roll after a colleague described it, so that session never watched the opening at all and lands on the reveal as extra views.

Name the evidence that would separate the stories before deciding which one you believe.

That last column is the part most teams skip, and skipping it is how you end up keeping the explanation that was already in the room — usually the one that justifies the edit somebody wanted to make.

Two things to hold back. Aggregate behavior cannot report an individual's feelings, so "the two people who rewatched it loved it" is not available from this data, even when the count is small enough that you think you know who they are. And "the version we already have" is not a neutral baseline. If the intro story feels truer because it arrives with a fix attached, that is a reason to distrust it rather than to act on it.

The cheapest test that would actually change your mind

Sort the follow-ups by cost, and take the cheapest one that resolves the uncertainty you have.

A bounded question. The audience is fourteen named people. One line — "Quick one: when you opened the sample, did you watch it straight through, or were you looking for something in particular?" — separates the content story from the viewing-task story more decisively than any further staring at the trace, because it asks about the task rather than the taste. Keep it non-leading. "Did the intro lose you?" is not a question; it is a request for agreement.

A declared revision. If you think the opening is the problem, change one thing the story predicts — the length of the intro, or where the first piece of the reveal sits — and show it to people who have not seen the other version. Be realistic about the workflow: you generally cannot send a second cut to the same recipient and read the result as a first impression, because they have already formed one. A revision test lives in the next batch, not the current one.

A measurement correction. If the problem is that nobody recorded who was told what, fix that prospectively rather than guessing retrospectively. One line per send — recipient, date, channel, any instruction or prior context — turns the next curve into something you can interpret. Least glamorous option, longest useful life.

Note which uncertainty none of these touch. If what you actually don't know is who received the link and what they were told, then no re-edit answers that, however elaborate it gets. An edit is an answer to an editorial question. It is not an answer to a distribution question.

When the recut goes to different people

Here is the trap, in the same invented production. A recut, version B, trims the sample to 2:40 and moves the reveal to 0:22. It goes to eleven recipients at two other companies, none of whom saw version A, in a note that says the reveal is near the top and they can jump to it. The trace is visibly gentler: about one session lost in the first half-minute, against four in version A; a bump at the reveal roughly one session wide; around half the sessions still live at the end.

The tempting sentence is: the recut fixed the drop. It is not available. Four things changed at once.

The recipients changed — different companies, different slates, different reasons for saying yes. The prior context changed: version B's recipients had already heard the show described on a call, so their first nine seconds were doing a completely different job from version A's first nine seconds. The instruction changed, and an explicit invitation to jump ahead will produce skipping in the record whether or not the material deserves it. And the window changed: six days of a smaller, more recent send against nine days of the first one.

Comparing two versions is still worth doing, sometimes. Keep the question, the window and the exposure conditions as close as the situation allows, keep the instruction identical, write down every departure, and treat whatever comes out as a reason to investigate or to leave something alone. Similar-looking conditions do not convert a casual before-and-after into a study, and two curves from two groups of eleven people are two anecdotes with graphs attached.

Leave the cut alone until you know who watched it

Back to the invented trace one last time. The steep fall is between 0:00 and 0:44. The plateau is the montage nobody would have defended in a meeting. The spike is two sessions wide, at the moment the machine plays.

Two readings survive contact with the evidence. Either the opening genuinely fails to earn the next thirty seconds, in which case the fix is editorial and specific — and the plateau may say less than it appears to, because of who was left watching by then. Or recipients opened a bare link, did their diligence inside the first minute, and skipped ahead to the part they had been told about. In the second case the material may be fine, and what needs fixing is the note that travels with the link.

The cheapest thing that separates those readings is one email to the people who already watched, sent before anyone opens the edit. If the answer comes back "I was looking for the reveal," the recut was about to solve a problem the recipient did not have. If it comes back "I stopped because the first minute never told me why to keep going," you have a target, and a version B comparison becomes worth running under conditions you can actually describe.

Until then, the curve has done its real job. It pointed at a moment. That is a coordinate, not a verdict, and the distance between the two is the whole of the discipline.

Frequently asked questions

Why can a small retention dip be misleading in a pitch sample?

At twelve viewing sessions, each session is worth a little over eight percentage points, so a five-point decline is a fraction of a person and below the instrument's resolution. The dip may be one or two named recipients rather than a trend.

What can a retention curve establish, and what does it leave unresolved?

It can point to where something happened in the video, such as a fall at 0:30 or a rise at 1:52. It cannot say why, and its shape alone cannot separate a viewer who stopped from one who skipped ahead.

Which conditions should be written down before interpreting the curve?

The article lists six: which upload and cut produced it; duration and observation window; the metric as the platform defines it; who received it, when, and what they were told; what recipients were asked to do; and how the link travelled. The last three often explain the curve.

What could produce a narrow bump at a reveal?

At least three rival explanations: recipients rewatched because the moment was valuable; they rewatched because it was unclear and they were trying to work out what happened; or the bump came from a forwarded timestamp or someone opening mid-roll. The curve alone does not choose among them.

What is the cheapest next step before ordering a recut?

Ask the named recipients one non-leading question—whether they watched straight through or were looking for something in particular. That separates the content story from the viewing-task story more decisively than staring at the trace. If the unknown is who received the link and what they were told, no re-edit answers it.

Further reading

• Can You Compare Streaming Views With TV Ratings in a Pitch? — Before comparing two versions or two retention numbers, read this on rebuilding each figure from numerator, denominator, population and window so unlike measurements are not treated as one.
• Is Retention Improving—or Are You Comparing Cohorts at Different Ages? — For the cohort arithmetic behind a retention curve, this explains aligning groups by elapsed age and keeping incomplete cells separate from zeroes so small-sample dips are not misread.