YouTube is a market for time, and everyone is measuring clicks

Every tool that measures YouTube measures views. A view is a click: it tells you someone started. It says nothing about whether they stayed.

We store how long every video runs — duration_sec, on 100% of the 12.4M rows we serve. That makes a second number computable: hours actually watched, once you discount a view for the part nobody sits through. It sounds like a minor variation on reach. It isn't. It's a different axis, and it disagrees with reach almost everywhere we look.

Everything below is measured against our production data across 15 niches, deduplicated per video, and restricted to channels that actually compete in the niche. Where a number rests on a modelling assumption rather than an observation, we say so.

How we discount a view

Before any table, the obvious objection: a view is not a completed view. Multiply views by duration and you credit a three-hour podcast with three full viewer-hours every time someone clicks it and leaves after ninety seconds. Do that and "attention" just becomes a leaderboard of who publishes the longest videos.

So we discount every view by how much of the video actually gets watched:

attention   = views × duration × retention(duration)
retention(L_minutes) = clamp( 0.79 − 0.124·ln(L),  0.15,  0.85 )

That curve gives roughly: a 3-minute video keeps 65%, ten minutes 50%, an hour 28%, three hours 15%. It averages ~43% across our corpus, which is inside YouTube's real 40–50% average-percentage-viewed range.

We did not measure this curve — nobody can. YouTube publishes no completion data. What we did is fix its shape from what's known publicly (retention falls with length, monotonically) and calibrate the constants against published studies — Backlinko's 1.3M-video analysis and a 10k-video/1,000-creator study. Our curve tracks their reported retention for strong performers at every length we can check: sub-1m 70–85%, under 5m 65–75%, 30–60m 25–35%, 60m+ 20–30%.

It matters enormously. Uncorrected, the >2h band looks like 34% of all attention; corrected, it's 17%. Half the number was an artifact of assuming people finish podcasts. Every figure below has the correction applied, and we flag the one place it doesn't matter.

Shorts win the clicks and lose the time

Group every video we track by length:

Read the last two columns against each other. Shorts are 34% of all reach and 2.2% of all time — they pull fifteen times more attention than they hold. At the other end, videos over two hours are 0.3% of the catalogue and 17% of every hour watched: over-represented roughly fiftyfold, but not, it turns out, the main event.

The main event is the middle. The 10–40 minute bands are 23.5% of videos and 47% of all watch-time — the deep review, the follow-along, the teardown. That's where the market actually is, and it's invisible if you rank by views.

Reach and attention aren't merely different. Across length, they're close to inversely distributed. A strategy optimised for one is actively wrong for the other.

Every niche is really five markets

Here's where it stops being trivia.

If you pool all lengths together and measure concentration, the number lies. It degenerates into "who makes the longest videos" — because the 40m+ bands carry 35% of all watch-time on 1.7% of the videos, they drag any aggregate toward their own structure.

Take Coffee & Espresso. Measured as one market, it looks like a tight oligopoly: roughly seven effective competitors. Split it by length:

Coffee is not a concentrated niche. It's two markets stacked: the short-form review market is a ~29-way free-for-all, and the long-form market is a 3-way lockup. The pooled "seven" was the long-form oligopoly bleeding through.

A creator who read the pooled number would conclude Coffee is closed and go elsewhere. The real answer is: short-form is wide open, long-form is owned by three players. Opposite conclusions, same data. You cannot get this from view counts, and you cannot get it from pooled attention either — it exists only once you band by length.

That generalises. For any niche we can now say: short-form is open (enter here); long-form is owned by N players (here's who).

We also checked whether banding is a real division or a convenient one. Per creator, the correlation between short-form success and long-form success is ~0.3 across every niche — the people winning short-form are, empirically, different people from the ones winning long-form. They're separate markets, not one market viewed at different zoom levels.

Two leaderboards, and most people only see one

If reach and attention disagree, then "who owns this niche" has two answers.

• Outdoor Boys — #2 in the entire corpus by watch-time. #1,132 by keyword footprint. It ranks for about 900 keywords and runs the second-largest attention economy we measure.
• Linus Tech Tips — #1 by watch-time, #8 by footprint.
• Formula 1 — #13 by watch-time, #8,846 by footprint. Fifty-seven keywords.

Footprint counts coverage: how much of the search surface you touch. Attention counts consumption: how much of the audience's time you hold. Both are real, both are useful, and they are not the same leaderboard. A competitive analysis that reports one and calls it "market share" is answering a question you didn't ask.

The niches themselves reorder too. Measured in watch-hours per day, they span a 32× range — from Gaming Peripherals at ~1.5M down to Home Office at ~48k — and the ranking doesn't match the reach ranking, because format mix converts reach into time at very different rates.

Wanting to buy and being able to buy are different things

One more divergence, because it's the one with money attached.

Two independent questions decide whether a niche's attention is worth anything commercially. Does the audience arrive wanting to buy? And are creators wired up to capture it — are there affiliate or commerce links on the videos holding the attention?

We can measure both, bottom-up. They disagree, and the gap is the interesting part:

Three regimes fall out. Where intent far exceeds wiring — Running at 80/26 — the audience shows up ready to spend and nobody has built the path. That's either untapped money or a category that monetises somewhere off-platform.

Where wiring far exceeds intent — Power Tools at 18/53 — it means our keyword signal is blind. Tool queries are model numbers, DeWalt DCD777, with none of the best / vs / review words that mark a shopping query. But the creators affiliate-link everything, because they know what the traffic is worth. Here the affiliate signal corrects the keyword signal.

And where the two agree — Gaming at 46/48 — you can trust the read.

What we're claiming, and what we're not

The band table, the concentration numbers, the reach-attention rank flips and the intent-wiring gaps are all measured. They come from data we already collect; no new pipeline was needed.

Two honest caveats.

The retention curve is calibrated, not measured. Its shape is defensible and it matches every published benchmark we could check, but the constants are a considered guess, because the ground truth is private to YouTube. So we tested how much our results depend on it — two answers, and they differ:

• Within a band it's a rounding error. Durations inside a band are similar, so the discount cancels out of the share maths: turn it off entirely and Coffee's short-form number moves 28.7 → 28.2, its long-form 3.1 → 3.2. The market-structure findings above don't rest on the curve at all — that's why we lead with them.
• Across bands it's load-bearing. It halves the >2h band, and at niche level it flips the #1 spot outright: raw, Golf leads on watch-time; corrected, Gaming does. Golf's lead was an artifact — 69% of its attention is 40m+ tour broadcasts that nobody finishes.

That's the honest split. The per-band conclusions are robust; the cross-band totals carry an assumption, and we say so every time we print one.

The wiring number is a floor. We detect affiliate and commerce domains in video descriptions. Link shorteners hide some, discount codes and sponsorships don't leave a link at all. Real commercial wiring is higher than what we count.

This is a research finding, not a product feature. The attention model isn't in our reports yet — it's the next thing we're building, on data that's already sitting in the warehouse.

Every figure above traces back to specific videos and searches in our dataset. You can poke at the same data yourself — the niches we cover, or check a channel, no signup.