AVD is not the same as watch time

AVD is not the same as watch time. Average view duration is mean watch time per view — total minutes divided by views — a typical stay. Watch time is the total minutes people spent. One is a mean. The other is a sum. Shortening a video will often raise AVD as a percentage of runtime and can still destroy watch time as a total. The number that pays on YouTube long-form is hours. The number that distributes a Short is closer to hold and complete. Pick the scoreboard the format uses before you cut for the mean.

A mean hides the shape

Two videos can share a 20-second AVD: one where everyone leaves at 20 seconds, and one where half leave at two seconds and half finish a minute. The retention curve tells those apart. Completion rate tells you who reached the end. AVD tells you the typical stay.

Studio still computes AVD on the engaged side of YouTube's 24 August 2026 split rather than on first-frame public views. Do not build a homemade AVD with public views in the denominator. After the cutover a swipe can write a public start and add almost nothing to minutes. That homemade mean will sag for a reason that is not the body. Engaged views are the stay pile. Use them.

Comparing AVD across lengths is how teams fool themselves. A 12-second clip with an 8-second AVD and a 12-minute video with a 2-minute AVD are not on the same scale. Compare within a length band, or read AVD as a percentage of runtime, or go back to the curve. Hook rate explains a low AVD that is really an opening cliff. Hold rate explains a low AVD that is really a body leak.

Rewatches pull the mean up. A looped Short can post an AVD above runtime. That is minutes from the same people, not a more typical first stay. Name it when you report.

Hours are the volume that pays

Watch time is everything between start and leave, summed. Two thousand people who each stay ten seconds write less watch time than two hundred who stay two minutes. A viral Short and a long-form explainer cannot be compared on views, and YPP still uses hours, not starts. Qualified watch hours are the subset that counts toward monetisation. Not every format, not every source of traffic, not every playback is in the numerator. Inauthentic or mass-produced uploads can be excluded. Mashable, 21 July 2026, recapped the buckets: generic near-duplicates; rage-bait without a story; synthetic personas as human authorities on healthcare, finance, law, or politics. AI stays monetisable when the work is original. Hours on a farm that later loses the programme are not hours you can take into next quarter.

A Short swiped at second one can now write a public view and still add almost nothing to watch time. If views rose and watch time did not, you bought starts on a hook that does not hold. That is the August 2026 story in one line.

RPM sits on top of this volume. The rate per thousand monetised playbacks will not save you if the thousands are first-frame starts. Lengthen a video only when the extra minutes have a job. Watch time does not reward padding. Unused minutes at the end of a curve are a slope, not hours.

Shorts RPM and long-form hours are different pools. Do not cut a 12-minute lesson to raise Shorts-style AVD% and then wonder where the qualified hours went.

The cut that raises AVD and loses the week

A creator sees AVD at 40% of a ten-minute file and cuts to five minutes. AVD% goes to 55%. Total watch time falls because the people who would have stayed through minute eight no longer have a minute eight, and the people who left at two still leave at two. The mean improved. The hours did not. YPP does not grade the mean.

A Shorts team sees AVD of four seconds on a 15-second clip and cuts to six seconds. Completion soars. Watch time per impression may rise or fall depending on whether the extra nine seconds were a leak or a proof beat. If they were the product demo, you raised the mean by deleting the ad. Saves will also fall if the useful specific lived in the deleted tail.

A thumbnail test lifts CTR and lifts public views. AVD falls because the new clickers are colder. Watch time may still rise because more people started. Report both. Killing the thumbnail because AVD dipped is reading a mean as a veto on reach.

Do not use AVD as a synonym for "people liked it." It is a typical stay. A 2-minute AVD on a 12-minute video can be a healthy session for that length band, or a disaster, depending on the curve. Open the graph.

Report the mean and the total on the same line

For long-form: watch time (and qualified hours when the question is YPP), AVD inside the length band, the curve. Never AVD alone.

For Shorts: engaged views versus public views, hold, complete, minutes. AVD as a share of runtime, compared to other Shorts, not to a 20-minute upload.

For ads: the buy's view definition, plus three-second hold. AVD on an ad is only as honest as the view definition in the denominator.

Captions still buy minutes on mute-first surfaces. Burn-in on Shorts and in-feed. Sidecar on the watch page so the player can keep people in a language they can read. The AI caption generator is the burn. Add captions to a video after you decide whether this file is a mean problem or a hours problem. Recutting length without recaptioning leaves cues that outlive the picture.

AVD is not watch time. Mean versus sum. Do not cut hours to decorate a percentage.

Keep the minutes that have a job, then read both meters, in Versely.

FAQ

Why did AVD go up after I shortened the video?

You removed tail minutes. Everyone who stayed now stays a larger share of a smaller file. Check total watch time. If hours fell, the mean is not a win.

Should I put public views in a homemade AVD?

No. Studio's AVD sits on engaged views after 24 August 2026. Public views include first-frame starts that never became a stay. You will understate the typical watch.

Which number should a sponsor see?

Hours and unique people, plus AVD labelled as a mean inside a length band. A single AVD without runtime and without a total is a statistic you can dress either way.