Open your retention curve and look for its shape, not its average. A cliff in the first seconds is the hook, a slow bleed is pacing, a step is one moment that broke belief, and a drop right after the pitch can be your buyers leaving to pay. Confirm each one by splitting the curve by device and by buyers against non-buyers before you rewrite a single line.
Read the shape, not the average
You cut the intro, re-uploaded, and the average watch time went up two points. Did the cut work? The average cannot say. A 41% average looks the same whether half your viewers watched most of the video or everyone left at the same moment, and those two videos need opposite fixes.
The retention curve shows where they left. In TrackPlay it is built every five seconds from the furthest point each session reached, so every point on the line is the share of viewers who got at least that far. Two numbers anchor it, both on the metrics catalogue: Hook Rate, the share still watching past five seconds, and Hold Rate, the share that reached the midpoint. How long should a VSL be? covers what those two say about length. This post is about the rest of the line.
First, what normal looks like. Wistia splits a typical curve into three parts: a drop at the start, a steady slide through the body, and a last dip at the end. A gentle downward slope is not a problem. Every video loses people. What you are hunting for is the place where the line stops sliding and starts falling.
Why a VSL curve is judged differently
Every guide to reading retention graphs was written for YouTube, and YouTube grades watch time because watch time is its product. A VSL's product is the order. The best viewer a VSL can have is the one who clicks the buy button at the pitch and never sees the rest of the video. On YouTube, that viewer is a drop. On a sales video, that viewer is the point.
So a VSL curve has to be read with the sales on it. When your cart posts each sale back to the play that earned it, any drop on the curve splits in two: the buyers who left, and the non-buyers who left. The same fall in the line means something very different depending on which group made it.
TrackPlay draws buyers as their own curve against everyone else, on the same timeline as the Pitch marker and the conversion density. That view needs one thing from you: a cart or affiliate postback, so the sale can find its play. Without it, you get the curve for everyone and nothing to split it by.
The six shapes
Most drop-off curves are made of a handful of shapes. Each one has a usual cause, a way to confirm it, and a fix. Confirm before you fix. A rewrite aimed at the wrong cause costs you a week of traffic and teaches you nothing.
| Shape | Likely cause | How to confirm | Fix |
|---|---|---|---|
| First-seconds cliff | Slow or self-focused opening, muted start, slow load | Hook Rate, Unmute Rate, device split, startup and buffering | Rewrite the first line to work with the sound off, add captions, handle the muted start |
| Slow bleed to the midpoint | A story or mechanism longer than the pain it carries | Hold Rate, overlay the previous cut | Trim the middle |
| Sudden step mid-video | A claim that broke belief, an early ask, a page change | Loop the range, check for a timed-event marker, buyers against non-buyers | Rewrite that section only |
| Step at the price | The price lands before the value | Pitch marker, conversion density, CTA shown against clicked | Build the value and the bridge before the number |
| Rewatch activity | A claim viewers want to check | Seeks in the event feed, rewatches in Hot Leads | Add proof at that claim |
| Post-pitch drop, then a plateau | Buyers leaving to pay, undecided viewers staying | Buyers-only curve, conversion density at the pitch | Leave the drop alone, work the plateau |
1. The first-seconds cliff
The line falls almost straight down in the opening seconds, then flattens. Some of that is normal on any video. Wistia measured an average loss of 4.9% in the first 2% of videos one to two minutes long, rising to 17.3% for videos five to ten minutes long. Longer videos lose more people at the door. A VSL is a long video.
Before you rewrite the hook, rule out the two causes that are not the script. Every major browser blocks autoplay with sound until the viewer interacts, so a VSL that autoplays usually starts muted. Check Unmute Rate: if few viewers turned the sound on, your hook was never heard. Then check the video's Quality tab for startup time and buffering, split by device. A video that takes too long to start on a phone loses the viewer before the first word.
If the delivery is fine and the sound came on, the opening is the problem. The fixes: write the first line so it works with the sound off, add captions so the muted opening still speaks, and use Smart Autoplay so a touch anywhere on the page turns the sound on without restarting the video. For the hook itself, How to structure a VSL covers what the opening has to do.
2. The slow bleed
No single fall, just a steady slide from the hook to the midpoint, steeper than the rest of the video. Hold Rate is low and no one moment is to blame. This is pacing. The story or the mechanism runs longer than the problem it is carrying, and viewers drift off one at a time.
To confirm, overlay the previous cut of the same video (the curve takes up to seven at once). If the older version held better through the same section, whatever you added there is the leak. The fix is trimming the middle, not the hook. AI Optimize reads the transcript against the real retention once a video passes about 300 plays and flags dead zones worth trimming, which gives you a shortlist to start from.
3. The step
The line runs level, then drops sharply at one moment, then runs level again at a lower height. Something happened at that second. The usual causes are a claim that broke belief, an ask that came too early, or a change on the page, like an order box appearing.
Drag across the step on the curve to zoom in and loop that exact section, and watch it as a stranger would. Then look for a timed-event marker at the same second. Every timed event you configured draws its own named marker on the curve, so a step that lines up with a page change is easy to spot. Finally, switch to buyers against non-buyers. If both groups drop there, the claim landed badly for everyone, and it is worth rewriting that section, only that section.
4. The step at the price
A drop when the price appears is expected. Some viewers were never going to pay, and this is where they find out. Do not judge this step on retention. Judge it on money.
The Pitch marker sits at the typical moment viewers reached the pitch. The conversion-density bars show where the buyers' watching peaked. CTA markers show how many viewers saw the button and how many clicked it. If plenty of viewers reach the price, see the button and do not click, the price arrived before the value did. The fix is upstream of the number: build the value and the bridge to the price first, so the number lands on a viewer who already wants the thing.
5. Rewatch bumps
On a Wistia or YouTube graph, rewatches show up as spikes where the line rises. A TrackPlay curve never rises, because it counts the furthest point each session reached, and scrubbing back does not move that. So rewatches live somewhere else: in the seek events of the User Events feed, and in Hot Leads, which scores each non-buyer partly on whether they rewatched.
A section viewers go back to is a claim they want to check. That is a good sign, and it tells you where proof pays. Put a demonstration, a number or a testimonial right at that claim.
6. The post-pitch drop and the plateau
The curve falls hard right after the pitch, then runs flat to the end. Every YouTube guide calls an end drop a failure. On a VSL it can be the best thing on the chart.
Switch the curve to buyers only. If the buyers' line falls at the same second, that drop is people clicking through to pay. Leave it alone. The viewers to work on are the ones on the plateau: they heard the offer, did not buy, and kept watching. Give them what the undecided need. An FAQ section, the guarantee spelled out, a second CTA. Continue Watching helps here too, offering a returning viewer Continue or Restart for 90 days, on any device.
Mobile against desktop: two curves, one video
A curve for all viewers blends two very different audiences. Split it by device and by orientation, and a mobile curve that falls earlier than desktop is common. The usual causes are not the script: a muted start that never turned into sound, a landscape cut played letterboxed on a portrait screen, or a slow start on a phone connection.
Check the Quality tab first. Then add captions, and upload a portrait version so phones get a video made for their screen instead of a letterboxed one. Split by traffic source as well. Cold traffic from an ad leaves earlier than an email list that already knows you, and a curve that mixes them will send you chasing the wrong section.
What is a good watch time?
The only honest answer: good for its length and its job. The public numbers come from general business video, not sales video.
- Wistia reports that videos under a minute average 52% of the video watched, and that the share drops 11% once videos pass 30 minutes.
- Vidyard's benchmark of 943,305 business videos found that 65% of viewers finish videos under a minute, against 20% for videos over 20 minutes.
No credible public benchmark exists for VSLs. Vendor pages quote ranges like "70 to 80% past the hook" with no dataset behind them. Treat them as guesses. The benchmark that counts is your own previous cut, on the same traffic, judged on Revenue per Play. A VSL that keeps fewer viewers and earns more per play is the better VSL.
Fix one drop at a time, then test it on sales
Change one section per edit. If you rewrite the hook, trim the middle and move the price in the same upload, the next curve will be different and you will not know why.
Then test the edit as a split test: clone the video, change the one section, and run both under the same embed, so the ad link never changes and each visitor sticks to one version. Set the goal to conversion rate, fed by the sale your cart posts back. The winner is declared only after the minimum sample and significance are met, so a lucky afternoon cannot pick it. Play counts are a vanity metric explains why the goal has to be the sale.
The one-paragraph version
Ignore the average and read the shape. A cliff at the start is the hook or the sound, so check Unmute Rate and delivery before you rewrite. A slow bleed is pacing, so trim the middle. A step is one moment, so loop it and rewrite that section only. A step at the price is judged on clicks, not retention. A drop after the pitch may be your buyers paying, so check the buyers-only curve and work the plateau. Split by device before blaming the script, fix one drop at a time, and let a split test on sales decide. If you want to read your own curve this way, the free plan gives you 3,000 plays a month, free forever, with no card.