How to Read an Attention Curve Like a Diagnostician
An attention curve isn't a grade — it's a map of where a video gains and loses its grip, second by second. Learn to read its shapes and every dip becomes an instruction for the next cut.
A single virality number tells you whether. An attention curve tells you where. That difference is everything, because "where" is the only thing you can act on. You can't fix a low score directly — there's no knob labeled "score." You can fix the specific second where the curve dips. The curve turns an opaque verdict into a list of edits.
But a curve only helps if you can read it. The shapes mean things, and once you know the vocabulary, a glance at the curve becomes a diagnosis.
What the curve actually is
An attention curve plots how much grip the video has over time, beat by beat. In Scrollproof, it's a fusion of per-frame visual saliency, motion, and audio energy — a model of how engaging the clip is at each moment, not a recording of real viewers (that's an important distinction, and we'll come back to it).
The height at any point is "how strongly is this moment pulling attention." The shape — where it rises, where it falls, how steeply — is where the information lives. You read shape, not height.
The shapes, and what they mean
A handful of curve shapes show up over and over. Learning them is most of the skill.
| Shape | What it looks like | What it's telling you |
|---|---|---|
| The cliff | A steep drop right at the start | The open isn't holding — a hook problem |
| The slow bleed | A clean start, then a gentle steady decline | The middle thins out — a hold/energy problem |
| The local dip | A sudden drop at one specific later moment | A dead spot — a slow transition, a tangent, a stall |
| The recovery | A dip followed by a climb back up | Something pulled attention back — a cut, a reveal, a turn |
| The plateau | A long flat stretch, neither rising nor falling | Steady but not building — fine, unless it's flat and low |
The cliff
A near-vertical drop in the first beat is the most common shape and the most fixable. It means the open failed to convert the stop into a stay. The instinct is to blame the whole video; the curve says look only at the first second. Start later, on a stronger frame. Land an audio event. Cut the runway.
The slow bleed
A clean start that decays gradually all the way down is a hold problem, not a hook problem. The open worked — people stayed past the first beat — but the video doesn't keep earning the next beat. The fix isn't a better open; it's more forward motion through the spine. Every beat needs to advance the thing the viewer is here for.
The local dip
This is the most actionable shape of all: a sudden drop at one specific timestamp while everything around it is fine. That's a dead spot, and it has a location. Go to that exact second. You'll almost always find a slow transition, a tangent, a moment where the energy or the information stalls. Cut it or compress it, and the dip flattens.
The recovery
A dip that climbs back up is a gift — it tells you what your levers are. Something at the bottom of that dip pulled attention back: a hard cut, a reveal, a change of pace, a new face. Whatever it was, it works in your hands. Note it, and use it earlier and more often.
Reading height vs. reading shape
A trap worth naming: people fixate on the height of the curve — "is it high?" — when the shape carries the diagnosis. A modestly-high curve with a clean shape is a healthier video than a tall curve with a brutal cliff. The cliff is where you're losing people; the height is just an average that hides it.
So read in this order:
- Where does it drop? Find the steepest losses first — that's where the leverage is.
- What's the shape of the drop? Cliff, bleed, or local dip — each points at a different fix.
- Where does it recover? Recoveries reveal your working levers. Steal from yourself.
- Only then, the overall height. As a sanity check, not the headline.
One honest caveat about model curves
A pre-publish attention curve is a model of engagement, not a recording of how real people watched — because no real people have watched yet. That's exactly what makes it useful (you get it before you publish) and exactly what bounds it (it's an estimate). It's very good at finding structural problems — a weak open, a dead spot, a flat stretch — because those are visible in the signal itself. It can't know that your topic will bore your specific audience, because that's about them, not the clip.
Treat the curve as a smoke detector, not a crystal ball. It reliably tells you where the structure is weak. It doesn't promise how the room will react. Both of those are true at once, and pretending otherwise is how trust gets spent.
The creators who get the most out of a curve aren't the ones chasing a tall line. They're the ones who treat every dip as a sentence — this is where you lost them, and this is probably why — and let the next cut answer it.
Stop guessing. Scan the clip.
Drop a short video and get Hook Strength, Hold Rate, a second-by-second attention curve, and a real attention heatmap — in about a minute. First scans are free.