AIflix started as the place we put video submissions that were too long for the live show but too good to lose. It has since turned into its own surface — a Netflix-style grid of AI-generated short films, music videos, animated sketches, and visual experiments. Because we run the player and the analytics on the same stack, we have a fairly clean view of how viewers actually behave. Some of what we see lines up with conventional short-form video wisdom. Some of it does not.
The TikTok playbook does not transfer cleanly
The default assumption a lot of AI video creators arrive with is the TikTok playbook: hook in the first second, no patience for set-up, cut every two beats, finish in under sixty seconds. That playbook was designed for a very specific viewing context — a phone in a thumb-scrolling feed where the next video is one swipe away.
AIflix is not that context. Viewers arrive on the page deliberately, scroll a grid, choose a thumbnail, and commit to clicking play. That single click is a much bigger commitment than a thumb scroll, and it changes everything downstream. Once the viewer has clicked, their patience window opens dramatically. Tracks that would die in a feed in two seconds will hold attention for the first thirty seconds of an AIflix play. The opening hook still matters, but the cliff is in a different place.
The cliff is around 90 seconds
Across our video plays, the steepest drop in retention sits around the 80–95 second mark. Plays that survive the first 90 seconds tend to keep going to a natural ending. Plays that don't, fall off sharply between 60 and 90 seconds. This shape is consistent across genre — animated, live-action style, music-video format, and narrative shorts all show the same general drop point.
Our read on this: 90 seconds is roughly the moment a viewer decides whether the piece is going somewhere or just looping its own aesthetic. AI video is uniquely vulnerable here because it is very easy to generate ninety seconds of beautiful nothing. Tools are good enough now that visual quality alone is not a moat. Structure is. Plays that survive the cliff almost always have an identifiable narrative move — a reveal, a turn, a punchline, an answered question — landing somewhere between the 45-second and 90-second marks. Plays that don't survive the cliff usually didn't have one.
Thumbnails do most of the work
On a feed-driven surface like TikTok, the first frame of the actual video is the thumbnail. On AIflix, the thumbnail is a separate surface — a still image the creator (or the platform) chose to sell the click. Tracks where the thumbnail differs meaningfully from the video's literal first frame consistently outperform tracks where the thumbnail is just a frame grab.
This is not a deep insight; it is the same lesson Netflix learned a decade ago when they started commissioning original key art for every title. But it is one a lot of AI video creators skip, because the easy path is to let the upload tool auto-pick a frame. The cost of skipping it is a measurably lower click-through.
Audio quality matters as much as visual quality
We see a small but real cluster of submissions where the visuals are striking and the audio is an afterthought — a placeholder score, a default text-to-speech narration, or an inconsistent mix between dialogue and music. Watch-time on these is consistently lower than on visually weaker submissions with stronger audio. We cannot prove the causal direction from our data alone, but the qualitative pattern is clear: viewers tolerate moderate visual weirdness if the audio is composed and balanced. They do not tolerate a striking visual with a hollow audio bed.
Series outperform one-offs, even when the one-off is better
Creators who upload three or four pieces under a recognisable identity — same title format, same recurring character, same visual language — accrue cumulative watch-time at a rate that outpaces creators who upload a single, highly-polished one-off. The library structure rewards the creator the audience can recognise. A viewer who liked one episode of a series will frequently come back for the next two; a viewer who liked a one-off has nowhere obvious to go next, and the library loses them.
This is, again, an unromantic finding. The tools encourage creators to chase the most ambitious single piece they can ship. The data says ship a smaller idea three times in a recognisable wrapper instead. We expect this to remain true even as the underlying generation tools get more capable, because it is a viewer behaviour finding, not a generation finding.
Where the format hits a wall
AI video tooling, in 2026, is still not great at sustained character continuity, dialogue-driven scenes that depend on precise lip-sync, or any storytelling beat that requires the viewer to track a specific object across cuts. Submissions that try to do those things land harder than they should. Submissions that play to the format's actual strengths — atmosphere, visual ideas, music-video pacing, surreal or dreamlike narrative — land more cleanly. This is going to keep changing as the underlying models change, but right now, the creators getting the most out of AIflix are the ones leaning into what the medium can do, not the ones fighting against what it cannot.
What we plan to add
We're working on per-piece chapter markers (so a viewer can see the structural shape of a longer piece before committing) and a creator-side retention chart (so creators can see exactly where their drop-off happens, instead of guessing). Both are in the backlog and neither has shipped yet. When they do, we'll write the follow-up: AIflix retention, six months in.