Raise information density to stop a slow bleed

Most retention problems are only visible after publishing. A slow bleed is the exception: it is fully present in the script, countable before a single frame is generated, and it is the only curve shape where the fix is arithmetic rather than taste.

The shape is unmistakable once you know it. No cliff at 0:20, no sharp step at 4:10, no obvious tangent to cut. Just a gentle, constant decline from beginning to end, ending somewhere respectable but never holding. Every diagnostic instinct fails on it because there is no timestamp to blame. That is the tell. A cliff means one moment failed; a bleed means every moment failed slightly, and the usual cause is that not enough was happening per unit of time.

Why a bleed is a density problem and a cliff is not

The retention curve shapes sort cleanly by which production decision caused them, and this is worth internalising because it tells you which department to send the fix to:

Shape	Signature	What actually failed	Where the fix lives
Cliff	Steep drop in the first 15–30s, then flat	The opening frame or line	Hook
Early plateau at a low %	Drops hard, then holds	Title and thumbnail promised something else	Packaging, not the edit
Mid-video cliff	Sharp drop at 40–70%	A tangent, an ad read, a format break	Cut that segment
Step-down	Small drops at regular intervals	Weak transitions between chapters	Bridge with a question
Slow bleed	Gentle constant decline	Too little new information per minute	The script

The four shapes above the bleed all have a location. You can point at them, watch the ten seconds before, and act. The bleed has none, which is why creators re-cut the intro they already re-cut, or blame the topic. Neither helps: the problem is distributed across the whole runtime.

Because the cause is distributed and informational, you can measure it on paper. That is not true of the other shapes.

The claims-per-minute score

Define a claim as one unit of information the viewer did not have ten seconds ago and could repeat to a friend afterwards. Four kinds count:

• Fact — a number, a name, a date, a specification. "25 frames per second, not 24."
• Instruction — a step the viewer can execute. "Duplicate one frame per second."
• Reversal — something that contradicts what the viewer walked in believing.
• Reveal — a visual that shows rather than tells, where the picture carries information the audio does not.

Four kinds do not count, and they are the bulk of what a bleeding script contains:

• Restatement of something already said
• Transitions and signposting ("so now let's look at…")
• Self-reference, channel business, and throat-clearing
• Hedging that adds no constraint ("it kind of depends on a lot of factors")

Scoring is mechanical:

1. Pull the transcript and break it into 30-second blocks. If you write your scripts before you shoot, score the script; if you don't, the transcript is the asset that lets you score after the fact.
2. Mark every claim with its letter: F, I, R, or V.
3. Count marks per minute. Write the number in the margin.
4. Put the per-minute counts next to the retention chart for the same video.

Step 4 is where it stops being a writing exercise. In a genuine bleed, the counts are flat and low across the whole body, and the curve is flat and declining across the same stretch. In a video with a mid-video cliff instead, you will see a single minute drop to zero while the rest holds — different diagnosis, different fix.

Here is what a scored six-minute explainer looks like when it bleeds:

Minute	Claims	Types	What was in it
0–1	5	F F R V I	Hook, the counter-intuitive premise, the setup shot
1–2	2	I I	Two steps, plus 40s of caveats
2–3	1	F	One number, restated three ways
3–4	2	I F	A step and its justification
4–5	1	V	A demo, narrated slowly
5–6	3	F I R	The actual payoff, arriving late

Minutes two through five are the bleed. The last minute proves the material existed — it was rationed. A strong idea stretched over a runtime it did not need is the usual form.

Finding your band

There is no published claims-per-minute dataset I would ask you to design around. The band is a number you compute from your own channel, and the procedure takes about an hour.

1. Pick your top quartile. Take your last twenty videos, rank them by average view duration relative to their own length, and take the top five.
2. Score all five. Same rubric, same 30-second blocks.
3. Take the median per-minute count across the body, excluding the first and last minute. Those two minutes obey different rules and will skew the average.
4. Do the same for your bottom five. The gap between the two medians is your band.

Once you have that, your edit target is simple: no minute in the body may score below your bottom-quartile median, and the body median should sit at or above your top-quartile median. Two consecutive minutes scoring zero is a structural defect regardless of what the medians say.

The one primary-source discipline worth borrowing is MrBeast's production document, leaked and widely reported in 2024. It assigns each stretch of runtime a job: minute one confirms the thumbnail's promise, minutes one to three deliver visible progression, later bands have their own work, with re-engagement beats scheduled rather than hoped for. That is density management by another name. The same document treats losing roughly 35% of the audience in the first 60 seconds as an above-average outcome — useful if you have been trying to eliminate early drop-off rather than beat your peers on it.

Editing toward the band

Once a minute is scored, the fixes are boring and effective, in this order:

Delete before you add. Every caveat, restatement, and signpost you remove raises the score without new material. A six-minute script edited to four minutes with the same claim count moves from 2.3 claims per minute to 3.5, and the viewer experiences that as pace. Publish at true length. Padding a six-minute idea to ten for a mid-roll threshold is a different shape: a drop around minute five or six, not a bleed.

Move the payoff earlier. If minute six carries three claims and minute three carries one, swap the structure. The reason to withhold a payoff is to set up a bigger one, not to fill time.

Convert telling into showing. A V-type claim carries information in the picture while the audio carries something else, which is the only honest way to raise density without speeding up delivery. Generated B-roll is genuinely useful here because you can produce the exact shot the sentence needs rather than the nearest stock approximation.

Do not confuse density with cut rate. They are independent axes. You can cut every two seconds and say nothing, which is a bleed with a headache attached. Cut rate is a platform-fit decision; density is a script decision. Get density right first, then set cut rate to the platform.

Re-check cheaply. The reason to fix a bleed in the editor rather than by re-generating is that a bleed is a subtraction problem, and subtraction is a timeline operation. The 480p preview pass costs nothing to run — it carries a short per-user cooldown, so you cannot spam it — and it is enough to confirm that a re-ordered cut still reads before you pay for the final export. Preview-first iteration matters more on a bleed than on any other shape, because you will run four or five passes before the pacing is right.

Then re-score. If the median moved and the next upload's curve flattens, you have the channel's floor. Log it. Score future scripts against it alongside your weekly numbers.

FAQ

How is this different from just "cut 20% of the video"?

Cutting 20% raises density only if you cut the right 20%. Scoring tells you which minutes are empty, so you cut from minute three rather than uniformly. It also catches the opposite failure — a video that is already dense and bleeding for a different reason, where cutting further would make it worse.

Does this apply to short-form?

Partly. Under 60 seconds the whole clip is effectively one density block, and the dominant signal is completion rather than curve shape. Score it, but expect the useful output to be "does this clip contain one claim or three" rather than a per-minute trend. The per-minute method earns its keep from about three minutes upward.

What if my score is high and the video still bleeds?

Then it is not a density problem, and you should stop treating it as one. Check two things: whether the packaging promised a different video, which produces an early plateau that can look like a bleed on a compressed chart, and whether the delivery is monotone. Constant pace of any kind causes habituation. Varying the rhythm — long holds punctuated by bursts — matters as much as the raw count.

Should I score every video?

No. Score your top five and bottom five once to establish the band, then score any new script that feels long while you are writing it. After ten or so passes you will estimate the count accurately by eye, which is the actual goal. The spreadsheet is training wheels.