Watch time per impression, as one number
Click-through and retention trade off. Combine them into watch time per impression, which YouTube's thumbnail test accumulates when it picks a winner.
A thumbnail that adds two points of click-through and dumps a third of those extra clickers in the first minute is not a win. It is a trade, and on YouTube the trade has a name: watch time per impression. Optimising CTR or retention on its own moves the wrong lever, because each one can rise while the product falls. The platform's own thumbnail test already resolves on the product. Your dashboard should too.
YouTube's Test and Compare help page states the winner is chosen by highest watch time, not by click-through rate. That sentence is the whole argument, written as a product decision. The rest of this post is the arithmetic that makes the same decision by hand, on videos that are not in a test, and on packaging choices the test will never see.
The composite, in one line
On a click-in surface (YouTube browse, search, suggested):
Watch time per impression = click-through rate × average view duration
Same quantity, from the two totals Studio already shows:
Watch time per impression = total watch time ÷ impressions
If watch time is in hours, convert first: hours × 3,600, then divide by impressions. The unit you want is seconds of watch time per impression. It is small, it is comparable across videos of different lengths, and it already includes the people who never clicked.
Worked example, numbers chosen so the multiplications close:
| Variant | Impressions | Clicks | CTR | Watch time | AVD | Seconds per impression |
|---|---|---|---|---|---|---|
| A, "clicky" thumbnail | 48,000 | 2,880 | 6.0% | 96.0 hours | 2:00 | 7.20 |
| B, quieter thumbnail | 48,000 | 2,112 | 4.4% | 117 h 20 m | 3:20 | 8.80 |
Check the two routes to the same cell. Variant A: 0.060 × 120 seconds = 7.20, and 96 × 3,600 / 48,000 = 7.20. Variant B: 0.044 × 200 seconds = 8.80, and (117 × 3,600 + 20 × 60) / 48,000 = 422,400 / 48,000 = 8.80. B accumulates 22% more watch time on the same impressions, with a CTR that looks like a loss in any report sorted by clicks.
That is the thumbnail test's decision, computed without the test. It is also why a CTR leaderboard and a Test and Compare result can point at different files. They are answering different questions.
On autoplay surfaces the click term drops out. The equivalent is closer to viewed-versus-swiped (or 3-second hold) × average view duration, still divided back through impressions. The algebra is the same. The packaging lever is the first frame rather than a thumbnail.
Why each half lies on its own
CTR without duration counts curiosity, including the curiosity you cannot pay. A title that over-promises inflates CTR and steepens the opening of the retention curve. The extra clickers are the people who were never going to stay. You paid for them with the next impression the system would have given a video that did not disappoint. Diagnosing a weak opening is the timestamp version of this; the packaging version is the same graph read against CTR. High CTR plus a cliff is oversell. Low CTR plus a cliff is a weak open. Only the second one is an intro rewrite.
Duration without CTR counts satisfaction among the people who already chose you. A 12-minute video with a 55% average view duration and a 1.8% CTR can be a better video than a sibling at 4.9% CTR and 28% held, and still lose the impression war because almost nobody entered. Watch time per impression catches that. Average view duration alone celebrates a video the browse page is not offering.
Completion rate has the same hole at a different length. A 45-second Short that 80% of viewers finish can still be a worse use of an impression than a 14-minute video that 35% of clickers hold to the end, once you multiply back through how many people started. Completions are a quality read inside a length band. They are not a packaging read.
The leaked MrBeast production document, widely authenticated and discussed in public, treats losing 21 million of 60 million viewers inside the first 60 seconds (35%) as an above-average result. Early drop-off is a rate to beat against peers, not a defect to drive to zero. That fact is about the shape of a successful video, and it is also a warning against worshipping retention percentages: a video can lose a third of its starters in a minute and still be the one the system should keep showing, if the remaining watch time per impression beats the alternative.
How to read it against a test, and against the rest of the channel
On any video currently in Test and Compare, stop using the CTR column as a preview of the winner. The test is accumulating watch time per variant. If the split was even, highest watch time and highest watch time per impression are the same ranking. If one variant got more impressions, divide each variant's watch time by its impressions before you treat the trophy as the composite. Then inspect why the winner won:
- Winner has higher CTR and similar duration. Packaging found more of the right people, or at least more people. Keep the winning image, and check the opening still matches it.
- Winner has lower CTR and higher duration. Packaging got more honest. Keep the winning image even if the CTR dashboard looks like a regression. This is the case the test exists to catch.
- No clear winner. Volume, or the variants were not different enough. The first-uploaded file stays. That is a sizing problem, not a metric problem.
On videos that are not in a test, compute seconds per impression once a week for anything still receiving a meaningful slice of browse or search traffic. Rank by that column, not by views. Views include subscriber delivery and leftover search. Seconds per impression is closer to "what did this impression buy."
A practical sheet, one row per video:
- Impressions (14 days, or the life of a test).
- Watch time in hours.
- Seconds per impression = watch time × 3,600 / impressions.
- CTR and AVD next to it, as diagnostics, not as scores.
- One-line cause if the composite moved: packaging, open, body, or length.
When the composite is down and CTR is up, you have an honesty problem in the thumbnail or title. When the composite is down and CTR is flat, you have a video problem; start with audience retention rather than a new still. When the composite is down and both halves look fine, check length. A padded runtime can hold percentage retention while destroying seconds per impression, because the extra minutes were never going to be watched.
Captions sit under this as a duration lever, not a click lever. They do not change the thumbnail. They change how much of the video a muted viewer can follow, which is why captions that actually hold watch time move the composite from the retention side. Treat them as a body fix.
The same composite is the reason to stop running packaging tests on videos that will not collect impressions. A test that never accumulates watch time cannot pick a watch-time winner. The broader A/B discipline still applies; the metric you pre-commit to, on YouTube browse, should be this one.
What you do differently on Monday
Stop reporting CTR and average view duration as rival KPIs. Report seconds of watch time per impression, then use CTR and AVD to explain the move.
Change packaging when the composite says the current still is expensive: high impressions, low seconds, high CTR. Change the video when the composite is low and CTR is already modest. Change nothing about the thumbnail because a CTR experiment "won" at 400 clicks.
For a working thumbnail set, keep the still that maximises seconds per impression against your last few siblings of the same format, even when it loses the CTR screenshot you would have posted in Discord. YouTube is already doing that in the test. Agreeing with it on the videos it is not testing is the whole habit.
The metrics that matter for brand video include this family of numbers. Watch time per impression is the one that stops CTR and retention from arguing.
FAQ
Is watch time per impression a YouTube-published metric?
No. Studio gives you the parts: impressions, watch time, CTR, average view duration. The product is one division. Test and Compare already decides thumbnail and title tests on watch time, which is the same product accumulated per variant.
Why would a lower-CTR thumbnail win?
Because the people it attracted stayed. Seconds per impression is CTR × duration. A 4.4% CTR with a 3:20 average view duration beats a 6.0% CTR with a 2:00 average on the same impressions. That is the 8.80 versus 7.20 case above, not a hypothetical.
Should I use this on Shorts?
Use the same algebra with the feed's entry rate in place of CTR: viewed versus swiped away, times average view duration, still per impression. Do not import long-form CTR into a surface that autoplays.