A host sent me a screenshot last month with one word attached: "help". It was a YouTube retention graph for a 74 minute episode. The line fell off a cliff in the first 40 seconds, wobbled along at about 45 percent for half an hour, spiked hard at 41 minutes, then slid into the floor. He wanted to know what he did wrong.
Here is the thing. He did almost nothing wrong. The graph was telling him something useful and he was reading it like a report card instead of a map.
I look at retention graphs most weeks now, because hosts send them to us before they book and after we deliver. I have opinions. The main one is this: most people read a retention graph as a verdict on the whole episode, when it is actually a list of timestamps you should go watch again. The number at the top is nearly useless. The shape is everything.
The average view duration number is the least useful thing on the screen
Everyone screenshots the big number. Average view duration, 14:32. Average percentage viewed, 22 percent. Then they compare it to some benchmark they read in a Reddit thread and decide their show is bad.
That number is an average of wildly different behaviours. It blends the person who bounced at four seconds because your thumbnail promised a different episode, the person who listened to 68 minutes while cooking, and the browse-page click who was never going to stay. Averaging those three tells you nothing you can act on.
What you can act on is the shape between the numbers. A graph is a sequence of decisions your listeners made. Every dip is somebody deciding you were not worth the next minute. Every flat stretch is a room full of people who stayed. Every bump upward is a rewatch, which means something in there was worth hearing twice.
So the first thing I do with a graph is stop looking at the y axis and start looking at the corners.
The first 30 seconds is a different graph, and it deserves its own zoom
Almost every retention curve has a near vertical drop at the start. That drop is normal. Some of it is people who clicked and realised it was not for them, and you cannot edit your way out of that.
But there is a version of the early drop that is your fault, and you can tell them apart. Zoom into the first 60 seconds. If the fall is a smooth slide, that is mostly audience mismatch, a packaging problem, not an edit problem. If the fall has a step in it, a sudden vertical cliff at, say, 0:18, go watch 0:18. Something happens there. Nine times out of ten it is the music sting, the "welcome back to another episode of", or the sponsor read that got moved to the front because someone said pre-roll converts better.
I have written before about how podcast intros lose listeners faster than any other single element, and the graph is where that shows up as a hard number rather than a hunch. A stepped cliff is a specific timestamp doing specific damage. Fix the timestamp.
A smooth early slide is a packaging problem. A stepped early cliff is an editing problem. They look similar on a zoomed out graph and they need completely opposite fixes.
Flat is the win condition, not high
New hosts want the line to be high. Experienced hosts want the line to be flat.
A graph that drops to 40 percent in the first minute and then holds 38 percent for the next hour is a good episode with a packaging problem. A graph that starts at 90 percent and bleeds steadily to 12 percent is a bad episode with great packaging. The second one feels better when you screenshot it and it is worse for your show.
Flat means the middle of your episode is not leaking. Flat means people who made the decision to stay were not given a reason to change their minds. Once the line is flat, the fix moves off the timeline and onto the title, the thumbnail, and the first 30 seconds. Our sister studio ClickTheory exists mostly because of that second half of the problem, and it is a different craft to editing. Eight thumbnails for $250 is a cheaper experiment than re-editing an episode that was already fine.
When I audit an episode, I write down two numbers: where the line settles after the intro, and where it is at the halfway mark. The gap between those two is the only retention stat I actually trust.
Dips are not always cuts, and this is where most people over-edit
Here is the mistake I see constantly. A host sees a dip at 23 minutes, goes to 23 minutes, finds a slow patch, and cuts it. Next episode, same dip, different minute. So they cut harder. Six episodes later the show is a tightly packed wall of talking with no breathing room and retention is worse than when they started.
Dips have causes, and only some of them are "this part is boring". In our sessions I look for four things first:
- A topic change with no signposting. The line dips because a listener could not tell if the new subject was for them. This is a chapter marker problem, not a cut problem.
- An audio step. Levels change, a guest goes quiet, a room tone shifts. People leave on discomfort long before they leave on boredom.
- A visual dead zone. On a multi-camera show, a two minute static wide is a dip machine. On a single camera show, the same two minutes are fine, because expectations differ.
- An actual dull stretch. This one is real, and it is rarer than the other three combined.
Human editors watching the whole episode is the only way I know to tell those apart, and it is why we do not touch silence detection tools. A silence-detection cut treats a thoughtful pause and a dead air fumble as the same event. Your audience does not. That difference is exactly the kind of thing that shows up as a soft dip you cannot explain from a spreadsheet.
The spike is the most valuable pixel on the chart
Rewatch spikes are gold and almost nobody mines them. When the line goes up, a meaningful number of people scrubbed backwards or replayed a section. They chose to hear it twice.
That is your clip. Not the bit you thought was the best moment, not the bit the guest was proudest of. The bit the audience voted for with their thumbs. I have argued that your best moment usually lands around minute 42, and the spikes on retention graphs are the closest thing to proof I have. People warm up. Guests relax. The genuinely good exchange happens long after everyone stopped performing.
Take the spike, back it up 20 to 40 seconds to find where the setup starts, and cut from there. That is your vertical. If you are choosing where a clip goes after that, the platform differences matter more than people expect, and I have gone through them in detail on shorts versus reels versus TikTok. We charge $25 per vertical clip and the spikes are where I would spend the first three.
If your graph has zero spikes across the whole hour, that is worth sitting with. It usually means the episode was pleasant and nothing in it was worth repeating.
Compare graphs against your own show, never against a benchmark
Retention benchmarks are close to meaningless across formats. A 12 minute solo explainer, a 90 minute interview, and a chatty two hander have completely different physics. Length alone swings average percentage viewed by a factor of four.
What is worth comparing is your own show against itself, same length, same format, five episodes deep. That is when the shape starts telling you about your choices instead of about your genre. Track three things per episode:
- Settle point after the intro
- Halfway retention
- Number of rewatch spikes
Five rows of that in a spreadsheet is more useful than any industry report. If your settle point climbs after you shortened your cold open, you learned something real about your show.
Also, and I will be blunt, do not check the graph on day one. Early data skews toward your subscribers, who behave nothing like the browse traffic that arrives later. Give it a week.
When the graph should not change your edit at all
There is a point where responding to retention data makes your show worse. If you are recutting an episode because of a two percent dip at minute 51, you have crossed it.
Retention data is a blunt instrument with a heavy delay. It tells you about broad shapes, not fine cuts. Editing your own show while staring at a graph is one of the fastest ways to lose your instincts, which is a big part of why I wrote about knowing when to stop editing your own show. A graph will always show you something to fix. That does not mean fixing it improves anything.
The other thing worth saying: retention problems often start at the recording, not the edit. A guest on bad audio from a laptop mic will bleed viewers no matter how good the cut is, and nothing in post fully rescues it. That is a capture decision, and it is why I keep going back and forth on Riverside versus Zoom for podcasts with hosts who are trying to save an hour of setup and losing twenty minutes of retention for it.
What I would actually do with your graph on Monday
Concretely, a 20 minute routine that beats most retention advice:
- Zoom into 0:00 to 1:00. Slide or step? Step means go watch that timestamp and cut it.
- Note where the line settles. Write it down.
- Note the halfway number. Write it down.
- Mark every upward bump. Those become clips this week.
- Look at the three biggest dips. Ask "audio, visual, signposting, or boring" before you cut a single frame.
- Do nothing else. Ship the next episode.
That is it. Six steps, no software, no benchmark comparison, no spiral.
If you would rather hand the whole loop to someone else, that is what we do. Human editors, first cut in 48 hours, revisions the next working day, and clips pulled from the moments your audience actually rewatched rather than the moments an algorithm guessed at. You can see the tiers and the packs on pricing, or just send me the episode and the graph at samy@pixelsamy.com and I will tell you what I see in it. I am in Dubai, so there is a decent chance you get a reply before you wake up.
Either way, stop treating the graph as a grade. It is a set of timestamps with opinions attached, and most of them are worth arguing with.