I have been uploading AI-generated Shorts to the same YouTube channel every day since 26 July 2026. Thirty-two uploads later the channel has gone from roughly 1,100 cumulative views to 4,375, and almost none of what moved that number was the thing I expected when I started.
What follows is what I actually changed, in the order the numbers told me to change it. No tool rankings — the generator matters far less than the four decisions below.
The specs that are not negotiable
YouTube treats a video as a Short when it is vertical and 3 minutes or under. In practice you want to be well inside that, and you want the frame right the first time, because re-uploading resets the video's entire performance history.
| Setting | What I use | Why |
|---|---|---|
| Aspect ratio | 9:16 | Anything else gets letterboxed and buried |
| Resolution | 1080x1920 | 720x1280 is accepted but visibly softer after compression |
| Length | 15 seconds | Long enough for one idea, short enough to loop |
| Audio | Always present | Silent Shorts stall — even ambient sound beats none |
| Captions | Burned in | Most views start muted |
The 15-second figure is not a style choice. A Short that loops is counted as watched again, and a 15-second clip with one clean idea loops far more often than a 40-second clip with three. If you are producing at any volume, standardising the length also removes a decision from every single video.
You must declare that it is AI — properly
This is the part people skip, and it is the one with actual consequences. A line of text in the description is not a disclosure. YouTube has a dedicated altered-content flag, and it is set on the upload itself.
If you upload through the API the field is status.containsSyntheticMedia, set to true on every AI-generated video. In the Studio interface it is the "Altered or synthetic content" question in the upload flow. Getting this wrong risks the video, and at scale it risks the channel.
It also costs you nothing in reach. Across our uploads the disclosed videos perform indistinguishably from anything else — the audience does not punish it. I have written up the full picture across every platform in the guide on AI video disclosure rules.
The first frame decides the video
Shorts autoplay, so there is no thumbnail doing the work for you. The first frame is the thumbnail, and it is on screen for the fraction of a second in which someone decides to keep watching or flick past.
The change that moved our numbers most was refusing to start on an establishing shot. A wide landscape that slowly pushes in reads as "nothing is happening yet" and gets flicked away before the interesting part arrives. Starting mid-action — the subject already doing the thing, at close range — was worth more than any prompt refinement.
The practical rule: if you can remove the first second and the video still makes sense, remove it. Then check the new first second is not also removable. Most AI clips can lose their opening beat entirely.
One shot beats five cuts
Early on I built Shorts as five timed cuts escalating to a reveal. It looked good in isolation and it consistently underperformed. The reason is specific to AI video: every cut is a fresh chance for the model to change your subject's face, clothing or scale, and viewers notice the inconsistency even when they cannot name it.
Moving to a single continuous shot — one camera moving through the whole fifteen seconds — removed the drift entirely and made the clips read as footage rather than as a sequence of generations. If you want the reasoning in full, it is the same principle behind keeping characters consistent across AI video.
Volume, and where it stops helping
Thirty-two uploads in ten days is about three a day, and three is deliberate. Pushing past three uploads to the same channel in one day measurably suppressed the other clips that day — the extra video did not add reach, it split it.
Three a day, spread across the day rather than posted in a block, has been the stable setting. Everything above that has cost us more than it returned.
What I would do on day one
- Fix the format before the content — 9:16, 1080x1920, 15 seconds, audio always, captions burned in. Decide once, never again.
- Set the synthetic-content flag on every upload. It is free and the downside of skipping it is not.
- Cut the first second. Then look at the new first second and consider cutting that too.
- One continuous shot. Cuts are where AI video falls apart.
- Three a day, spread out. More is not more.
The short version
The generator you pick is close to irrelevant for Shorts. Vertical at 1080x1920, fifteen seconds, sound on, captions burned in, the synthetic-content flag set, opening mid-action, in one unbroken shot, three times a day. Our channel went from about 1,100 to 4,375 cumulative views in five days on exactly that recipe, and every improvement in it came from removing something rather than adding it.
One pricing trap worth knowing before you subscribe anywhere: Captions sells two different products under one name, and only the higher tier includes any AI generation at all. Which features sit behind which price is in Captions AI pricing.
Looking for something else? Browse all 72 AI video guides in one list.
If you want to see the actual uploads behind those numbers — which ones flopped, and what the first frame looked like when they did — that is what I share inside the AI Video Generator community on Skool. Bring your channel and we will look at your first frames together.
Share:
Higgsfield Pricing in 2026: What a Video Actually Costs
How Many AI Videos Should You Post Per Day? The Answer Is Three