The single most common mistake we see in generated short-form video is treating 9:16 like a 16:9 frame turned sideways. It is not. Reels, TikTok, and Shorts have their own creative grammar, and ignoring it is the cheapest way to torch your ad budget. Here is what we have learned about the vertical format from watching thousands of generated clips perform.

Some context first. Three years ago the dominant ad video format was 16:9 because that is what YouTube pre-roll demanded and YouTube was where the budget went. Today more than 60% of ad spend on Meta moves through Reels placements, TikTok demands native 9:16, and YouTube Shorts now drives the bulk of new YouTube ad inventory. If your creative still gets produced as 16:9 with a "vertical crop pass" added at the end, you are several years behind.

Why vertical is different

Three structural differences change everything about how you compose a vertical ad.

The viewer holds the screen. Phone screens are six inches away from the viewer's face and the viewer is alone with the screen. The intimacy is closer to a one-on-one conversation than a billboard. Wide establishing shots make no sense in this format. The viewer wants to be talked to, not panoramically impressed.

Vertical eats text differently. A horizontal frame has room for two lines of caption text in the safe area. A vertical frame has room for one line, often less, because the top and bottom of the canvas are reserved by the UI of every platform (caption bars, like buttons, user avatars). Plan for the chrome.

Vertical has no horizon. Most cinematic composition rules assume a horizon line and use it as the eye's anchor. Verticals do not get one. Composition is built on vertical lines and centered subjects instead. Trying to apply the rule-of-thirds the way you would in a 16:9 frame produces awkward, off-center clips that feel mistaken rather than intentional.

The hook window is 1.7 seconds, not 3

The conventional wisdom is "you have 3 seconds to hook a viewer." That number came from horizontal pre-roll. The vertical equivalent is 1.7 seconds. We have measured it across hundreds of campaigns. After 1.7 seconds, the swipe rate stabilizes — anyone who is going to leave has left.

What this means practically: your value proposition needs to be visually obvious in the first half-second, your pattern interrupt needs to land before the one-second mark, and your hook needs to resolve into something interesting by the 1.7-second mark. Three full seconds is a lifetime. You probably have closer to two.

The implication for prompt-writing: when you brief a model for a vertical ad, the opening frame matters disproportionately. We tell teams to write the prompt for the first frame, then write the prompt for the rest of the clip as a second pass. Generating an 8-second clip with a generic "the product being used" prompt produces an opening half-second that looks like every other ad in the feed.

Composition rules that actually work in 9:16

After cataloging thousands of high-performing generated verticals, four composition patterns repeatedly outperform everything else:

1. Centered subject, vertical motion. One subject, dead center, motion that travels up or down rather than across. Why: the eye finds the subject instantly and the vertical motion uses the format's natural axis. Used in roughly 40% of top-performing UGC-style ads.
2. Stacked composition. Three visual elements stacked vertically — header product shot at top, demo in middle, CTA at bottom. Why: gives the eye something to discover at each height. Especially good for product-feature ads.
3. Talking-head close-up. A person's face filling the top two-thirds of the frame with caption space below. Why: faces are the highest-engagement visual element on the platform, period. The format was practically designed for this composition.
4. The hand-action shot. A close-up of hands doing something with the product. Why: it implies the viewer's own perspective and pulls them into the action without the uncanny-valley problem of CGI human faces.

Compositions to avoid: anything with a horizontal element as the focal point (a car driving across the frame, two people standing side-by-side), and anything that requires the viewer to read text larger than five words to make sense of.

Audio is half the work, and most teams ignore it

About 80% of vertical ad watches happen with sound on by default. Compare that to 15-20% for in-feed horizontal ads. The implication: your ad's audio is doing as much performance work as your video.

The generative models we recommend for ad video do not yet produce native audio that works for ads. Veo's audio output is improving but still feels uncanny. Our standard workflow is: generate the video, then overlay one of three audio assets:

1. A 12-15 BPM-matched track from your music library (the platform-native track libraries are now excellent and royalty-free).
2. A voice-over written specifically for the visual, recorded in 30 seconds on a phone (yes, a phone — the production value of "raw" voice is now a feature, not a flaw).
3. Ambient sound that matches the visual environment (kitchen sounds for a kitchen ad, etc.). Adds realism and lifts engagement noticeably.

Skip generic background music. The platform's algorithms have learned to recognize stock music as a signal of low-effort content and will demote your ad accordingly. Whatever you do for audio, do it deliberately.

The first frame should give away the genre

Viewers categorize videos in the first half-second. They are looking for "ad," "UGC," "how-to," "comedy," "news," or one of about a dozen other genres. Once they have categorized, they decide whether to keep watching. If they cannot categorize, they swipe.

The frames that fail the categorization test are usually trying to be coy — withholding the genre to "build intrigue." That worked in 30-second TV spots. It does not work in two seconds of feed time. Tell the viewer immediately what kind of thing they are about to watch. The interest comes from the specific take within the genre, not from suspense about which genre it is.

This means: if you are running a product ad, the product should be visible in the first frame. If you are running a story-format ad, the protagonist should be visible. If you are running a UGC-style ad, the person should be talking directly to camera. Pick a genre, signal it instantly, then earn the next four seconds with something specific.

Captions are the new headline

Even with audio-on rates as high as they are, hardcoded captions consistently lift performance by 15-25% across our test data. The reason is not accessibility. It is bandwidth. The viewer's eyes are processing the visual and the captions in parallel — that is twice the information per second. The captions get to deliver the value proposition while the visual delivers the emotional hook.

Caption craft matters. Generic captions ("Check this out!") do worse than blank captions, because they consume attention without paying back. Specific captions ("This is the only serum I've kept buying for two years") earn the screen real estate. Write captions like headlines, not like subtitles.

One technical note: every platform now auto-captions, but auto-captions look generic and read awkwardly. Hardcoded captions in your visual style outperform auto-captions by a noticeable margin. They are also the easiest single thing to improve about your existing ad library.

Aspect ratios beyond 9:16

A quick note on adjacent ratios: 4:5 (used by Instagram feed posts) and 1:1 (used everywhere) are not the same creative problem as 9:16. They want different compositions. 4:5 in particular is a trap — many teams ship a 4:5 ad that is actually a center-cropped 1:1, which loses the format's slight vertical advantage.

Our recommendation: generate three explicit versions of every campaign — 1:1, 4:5, 9:16 — each with composition decisions appropriate to the format. Generating once and cropping produces three mediocre ads instead of three sharp ones. Modern platforms support per-format generation cheaply enough that there is no reason to cut corners here.

How to test if your vertical creative is working

The diagnostic metric is the swipe-up rate at the 1.5-second mark. If more than 60% of viewers are still watching at 1.5 seconds, your hook is doing its job. If fewer than 50%, the hook is failing and no amount of clever middle-of-clip work will save it. Iterate on the opening, not the body.

Second diagnostic: the 6-second hold rate. If viewers are still watching at 6 seconds, you have earned the right to ask for a click. If hold rate cliffs between 1.5 and 6 seconds, your hook is working but the body is not paying it off. Fix the middle.

Most importantly, look at the swipe rate before looking at the click rate. A 9:16 ad with a 1% CTR and a strong hold rate is a real creative success that just needs a better offer. The same CTR with a weak hold rate is a creative failure dressed up as a conversion problem. Different fixes apply.

The summary you can put on a wall

If you remember nothing else from this post: 9:16 is not a smaller 16:9. The hook is shorter. The composition is centered. Audio is on by default. Captions are mandatory. Genre signaling happens in the first frame. The viewer is alone with the screen and wants to be talked to, not impressed.

Generate every campaign three times — once for each major aspect ratio — and write the prompt for the format, not against it. The teams that do this consistently are the ones whose vertical ads stop looking like recycled YouTube ads and start looking like they belong in the feed. That is the bar, and it is fully achievable with the current generation of models if you respect the format.