AI video ads: how to make vertical ads and UGC-style content that stop the scroll

What makes a vertical video ad stop the scroll

In a vertical feed, your ad competes with the next swipe, and the decision happens in the first moment. Whatever the platform, the ads that hold attention tend to share four traits:

1. Motion from frame one. No logo sting, no slow fade in. The first frame should already be mid-action.
2. One idea per clip. One subject, one action, one feeling. A five-second clip with two ideas delivers neither.
3. Readable with the sound off. Many people scroll muted. The picture and the on-screen text must carry the message on their own.
4. Native to the feed. Shot (or generated) for 9:16, with the subject in the center, looking like content rather than a cropped TV spot.

This clip does all four. From the very first frame the dog is already running at the camera through a puddle, with water arcing around it in slow motion. The yellow raincoat is the only saturated color in a grey street, so the eye finds it even at thumbnail size. Five seconds, one action, no cut. It cost 57 credits.

How to write a hook prompt for AI video

A hook prompt is short on purpose. Every word should describe something visible in the first second.

• Subject plus one bold visual tag. A familiar subject with one detail that pops: a color, a costume, an odd scale. The tag is what makes it recognizable at a glance.
• An action already in progress. "Splashing", not "starts to run". Verbs in the -ing form tend to give you motion from the first frame.
• Setting in a few words. Enough to explain the light and the backdrop, no more.
• Motion and format. Slow motion stretches the satisfying part. "Vertical video" helps composition, but also set the aspect ratio to 9:16 so the model composes for the frame rather than you cropping later.

Most short-form hooks fall into a handful of patterns. Our four test clips cover the first four:

Generate hook variants at volume on cheap models

You can't predict which hook will win, so the economics that matter are cost per variant and time per variant. All four of these clips were generated from text on PixVerse V6, at 9:16, five seconds each, for 57 credits apiece.

Watch each one through before you judge it. The dachshund and honey clips hold together for all five seconds. The jellyfish clip starts as a field of drifting jellies and ends on a close-up as the camera pushes in, which gives it a natural payoff. The robot clip has a continuity slip: halfway through, the watering can moves from one hand to the other and the arms rearrange. In a feed, few people would notice. In a product ad, a similar glitch on your product would be a problem.

Here is how that compares with the premium clips in our gallery:

One 305-credit Seedance 2.0 Fast clip costs as much as 5 of these 57-credit hooks, and a batch of ten hooks costs 570 credits ($5.70). That is the core strategy: test ideas on inexpensive models, then spend premium credits on the winning idea when you need better physics, longer shots, lip sync or native audio (Veo 3.1, for example, generates sound with the picture). Rankings move fast: at the time of writing, independent arenas rank several newer video models above the ones used here, so check the model catalog for what is live, and see our lesson on choosing a video model for how to compare them.

The testing loop: ten hooks, ship, read, iterate

Generating variants is only useful if you learn from them. The loop below keeps every round a fair test.

1. 01

   Write one template with variables

   Turn your best hook prompt into a template, with the parts you want to test as variables. Keep everything else fixed.

2. 02

   Generate ten variants

   Fill the variables ten ways and render them all on the same model, aspect ratio and length. Discard any with visible glitches.

3. 03

   Give every hook the same body

   The AI clip is the opening seconds. Follow it with the same product shot, message and offer in every version, so the hook is the only difference.

4. 04

   Ship them as separate ads

   Same audience, same budget, same placement. Let them run long enough to get a fair read.

5. 05

   Read hook rate, then hold rate

   Hook rate: the share of impressions that watch past the opening seconds. Hold rate: of those, how many keep watching. Then check clicks and cost per result. A hook that stops thumbs but loses viewers right after is mismatched with the body.

6. 06

   Keep the winner, change one variable

   If the dachshund beat the robot, the next round tests dogs in different outfits or weather, not ten new animals.

PromptLab's Consistency runs are built for this step: you write a template with variables, a drafter proposes values, and the whole matrix (up to 10 variants) renders in one run with pass/fail scoring, so glitchy variants are filtered before they reach an ad account.

Design for sound off: captions, safe zones and 9:16

A clip is not an ad yet. The finishing steps are where most of the clarity comes from.

• Generate at 9:16. Cropping a 16:9 clip to vertical loses most of the frame and often the subject. Set the aspect ratio in the model settings, not just in the prompt.
• Respect the safe zones. On TikTok, Reels and Shorts, the caption, username and action buttons cover the bottom of the frame and the right edge. Keep faces, products and text in the central area, and check each platform's current safe-zone template before export.
• Add on-screen text in the edit. Video models still misspell, and ad copy, prices and legal lines must be exact. Generate the picture, then set the type in your editor.
• Caption every spoken word. Burned-in captions make voiceover readable with the sound off.
• Make frame one a thumbnail. Pause the clip on its first frame. If that still doesn't say what the ad is about, change the hook.
• Add sound anyway. The clips on this page play without sound. For viewers with the sound on, add music, effects or a voiceover. PromptLab Flows has Text to Speech and Music nodes for a voiceover and a music bed from a brief.

How to make AI UGC ads without fake testimonials

"UGC-style" describes a look: a person talking to a phone camera, handheld, natural light, casual delivery. AI can produce that look, and as a format it is fine. The line is crossed when the ad presents a synthetic person as a real customer with a real experience.

In the US, the FTC's rule on consumer reviews and testimonials took effect in October 2024. It prohibits creating or disseminating testimonials that misrepresent that they come from a person who doesn't exist, or from someone without actual experience of the product, and the FTC's announcement names AI-generated fake reviews as an example. The FTC can seek civil penalties against knowing violators. Its Endorsement Guides separately require disclosing material connections, such as paying a creator. Other countries have their own consumer-protection rules; this is not legal advice.

Platforms add their own layer. TikTok asks creators to label realistic AI-generated content and applies an AI label automatically to uploads carrying C2PA Content Credentials. YouTube requires creators to disclose realistic altered or synthetic content. Meta shows an AI disclosure on ads made or significantly edited with its generative tools, and detects some third-party AI content from C2PA metadata. Ad rules tightened during 2026 and differ by region, so read each platform's current ad policy before launch.

If you want a talking presenter, build an obvious brand character and own it. PromptLab Flows has a Lipsync node to drive a face with an audio track, and a "Talking-head onboarding" starter template that shows the pattern.

Keep the product accurate

Text-to-video draws a product that looks like yours. For an ad, it has to be yours. The reliable route is image-to-video: start from a real packshot (or an approved still of your real product), and write a motion prompt that describes only what moves.

Start frame

→

slow rotation, 57 credits

Image promptFlux 2 Pro · 8 creditsCopy

Tortoiseshell aviator sunglasses floating above rippled desert sand at dusk, the lenses reflecting a pink and orange sky, long shadow, luxury eyewear campaign, hyper-detailed

Motion promptPixVerse V6 · 5 s clip · 57 creditsCopy

The sunglasses rotate slowly in mid-air, fine sand drifting below them, the sky colours shifting from pink to orange, product commercial

This product clip also ran on PixVerse V6 for 57 credits, animated from the start frame shown with it. It is a good reminder to review every frame: as the sunglasses turn, the shape of the bridge and the temple arms shifts from one moment to the next. That is acceptable for a mood clip and not for a product ad. Fast rotations and complex physics are where products deform; slow moves and simple light changes are safer. Re-roll, try a stronger model, or cut around the frames that drift.

The same honesty rules as product photography apply: the ad can only show features and results the product really has. Our AI product photography guide covers reference images, and the image-to-video lesson covers motion prompts shot by shot.

Production checklist for AI video ads

1. Pick one idea and one hook pattern per ad. Check what is trending in your niche first.
2. Write a hook template with variables; generate ten variants at 9:16 on an inexpensive model.
3. Watch every variant to the end. Discard glitches, warped products and invented text or logos.
4. Animate your real product from a packshot; never let text-to-video invent it.
5. Add on-screen text, captions, sound and the offer in the edit, inside the safe zones.
6. No synthetic customers, invented reviews or cloned real people. Disclose paid partnerships.
7. Turn on AI labels where platforms require them for realistic content.
8. Ship variants with an identical body and budget; read hook rate, hold rate and cost per result.
9. Keep the winner, change one variable, repeat. Remake the winning idea on a premium model if it earns it.

Two PromptLab features help at the edges of this loop. Trends pulls TikTok and Instagram posts for the hashtags you choose, which is a quick way to see which formats are working before you write hooks. Social publishing posts finished clips to your connected accounts. Neither replaces the testing; they shorten the trip between idea and result.

Questions people ask