Testing AI Video Ads: Angles, Hooks and What to Measure

The pitch for AI video ads usually stops at "you can make more of them, faster and cheaper." That is true, and on its own it is the fastest way I know to waste an ad budget. Testing AI video ads well is not about producing more creative. It is about producing creative that asks better questions, so each round of spend teaches you something you can build on.

I run the studio, so I sit on the commercial side of these conversations: what a client should spend, what they should expect to learn, and when to stop. This post is the framework I walk founders and marketing leads through. It covers what AI actually changes about testing, why angles matter more than variations, how to structure a batch so the results read cleanly, and what to do with the numbers when they come back. None of it requires a big budget. It does require discipline.

What AI changes about ad testing, and what it doesn't

Ad creative wears out. A winning video scales, frequency climbs, results slide, and the account needs something new. The traditional replacement means briefing, shooting and editing, and it often arrives after the damage is done.

AI changes the production maths. Creators, settings, product scenes and animation can be generated instead of filmed, so a batch of ten different ideas costs a fraction of ten shoots. That makes it realistic to test ideas you would never have gambled a shoot on.

What it does not change is how learning works. A test only teaches you something if the ads are meaningfully different and the results can be traced back to that difference. Fifty near-identical ads teach you roughly as much as one. And audiences notice lazy AI: one visual trick, generic avatars and no real edit get scrolled past quickly.

Angles first, variations second

This is the idea I push hardest, so it is worth defining terms.

An angle is a different argument for buying. "It saves you two hours a day" is one angle. "Your competitor is already using it" is another. "Here is the objection everyone has, answered" is a third. A variation is the same argument dressed differently: another creator, another background, another colour on the caption.

Early in testing, variations are mostly noise. If you test five versions of one angle and one does slightly better, you have learned something small about styling. If you test five genuinely different angles and one clearly wins, you have learned what your market responds to, and that informs your landing page, your emails and your sales calls, not only your ads.

Some angle types that tend to produce useful contrast:

• Problem-first, opening on the pain before the product appears.
• Outcome-first, opening on the end result.
• Objection, naming and answering the reason people hesitate.
• Comparison, against the old way of doing it, not a named competitor.
• Curiosity, a question or an unexpected image that demands the next second.
• Offer, where the deal itself is the message.

What to test at each stage

Not everything should be tested at once. Here is how I think about the order, and what each level actually tells you.

Test the big questions first, when differences are large and easy to read, and save fine-tuning for the ads that have earned it.

Picking the right format for each angle

One of the genuine advantages of AI production is that you are not locked into one format for a whole batch. Each angle can be made in the style that suits it.

An emotional, relatable angle often works best as an AI UGC video with a creator talking to camera. A mechanism, how the product works inside something, usually works better as AI animation. A premium physical product often needs a cinematic product shot. A sharp claim can be carried by bold kinetic typography alone.

Mixing formats also helps with fatigue. An account running the same visual style in every ad burns through its audience's attention faster than one that rotates looks. Our post on creative fatigue in paid social goes further into why that happens and how to plan around it.

A scenario: the DTC brand with a flat account

Picture a Shopify brand selling a posture corrector, spending steadily on Meta and TikTok. Their best ad is a creator-style video that has been running for months. Results are sliding, and they have been making small tweaks to it: new music, a different first line, another creator. Nothing has moved the needle.

The problem is that every new ad is a variation of the same angle. Here is what I would suggest instead. A batch of around eight angles: the desk worker's afternoon back ache (problem-first), someone standing straight in a meeting (outcome-first), "does it show under a shirt?" (objection), the cost of a physio course versus the product (comparison), and so on. Each is made in the format that suits it, one as AI UGC, one as a short animation of the spine, one as a clean product film, and cut in 6 and 15 second versions for each placement.

When results come back, they are likely to show that one or two angles have life in them. The next batch then iterates on those: new hooks on the winning angle, a new format for the runner-up. Each batch builds on the last rather than starting again. For the specific e-commerce side of this, our e-commerce industry page covers how product video, UGC and ads fit together for DTC brands.

Structuring a batch so results read cleanly

A test is only as good as your ability to read it afterwards. Three things make that possible.

1. Name every file so it explains itself. We use a convention like brand_angle03_hook02_ugc_9x16_v1: angle, hook, format and ratio in every filename. When the results export from your ad account, you can sort by angle or format without opening a single video.

2. Keep a variant map. A simple sheet listing every ad and what makes it different. Six weeks later, nobody remembers which hook was which.

3. Cut hooks as separate modules. If the opening is its own piece, you can pair a winning hook with a new body, or test three hooks against one proven body, without rebuilding the ad.

What to measure, and what to ignore

For each ad, the numbers that matter most are simple: spend, hold rate (how many people keep watching past the opening), and cost per result, whatever your result is. Hold rate tells you whether the hook works. Cost per result tells you whether the whole ad works. An ad with a strong hold rate and a poor cost per result usually has a good opening and a weak body or offer, which is useful information in itself.

What I would ignore early on are small differences between ads that have barely spent. Give each one enough budget and time for the result to mean something before calling a winner. How much that is depends on your price point and your account, and your media buyer is better placed than any creative studio to judge it.

Who this is for, and who it isn't

This approach suits brands spending on paid social whose creative cannot keep up with fatigue, SaaS and app teams wanting to test messages before committing to a full shoot, performance agencies that need production capacity behind their own strategists, and founders about to spend their first real ad budget who want to learn fast.

It is less suited to a brand spending so little that no ad can gather meaningful data; you would be better off with fewer, stronger ads. It is also not a replacement for someone running the account. We produce the creative and hand it over, organised for whoever buys the media. If your ads are built mostly from your own product UI or screen recordings, our ad creative systems service uses the same test-first structure with motion design instead of generated footage.

How we produce AI video ads at Degird

Our AI video ads service runs as a loop rather than a one-off. It starts with a teardown of what you sell, what your ads have already said and how they performed, and what runs in your category. The output is a numbered angle list with a format and hook for each, approved before anything is made.

Next, product plates, any AI creators and the style frame are locked, and one ad is built end to end to set pacing, captions and graphics. You sign it off, then the rest of the batch is generated, selected at full size and edited against that system, with hooks cut as separate modules. Every ad is reframed for 9:16, 4:5, 1:1 and 16:9 where needed, checked against platform overlays, captioned, named to the convention and delivered with the variant map. A typical batch is five to ten distinct angles in 6, 15 and 30 second cuts, plus hook-only variants.

Then you send back spend, hold rate and cost per result per ad, and the next batch extends what won. Because the product plates, creators and style are reused, later batches are faster and cheaper than the first.

AI video ads are quoted per batch against the approved angle list, so the number of ads, lengths, ratios and formats are fixed before production. Accounts refreshing creative every week are usually better served by the monthly retainer, working through angles one request at a time; both options are on our pricing page.

Frequently asked questions

How many AI video ad variations should I test at once?

Enough that each one tests a different idea. For most accounts that means five to ten distinct angles per batch. More versions of the same angle rarely teach you much; more genuinely different angles do. Once a winner appears, the next batch iterates on its hook and format.

Do AI video ads perform better than traditional video ads?

There is no honest general answer, and I would be wary of anyone who gives one with a percentage attached. The format matters less than the angle and the craft. What AI reliably changes is cost and speed, which lets you test more ideas for the same budget and find the angle that works sooner.

What metrics should I use to judge an AI video ad test?

Spend, hold rate and cost per result for each ad. Hold rate shows whether the hook stops people; cost per result shows whether the whole ad does its job. Read them together, and avoid calling winners on ads that have barely spent.

Will Meta and TikTok approve AI-generated video ads?

AI-generated ads go through the same review as any other creative and are held to the same policies on claims, before-and-after results and impersonation. Both platforms also ask for disclosure of realistic AI-generated content. We produce within those rules and flag any claim you will need to substantiate.

Can AI video ads be made from our existing product photos?

Usually, yes. Clean product photos, ideally front, side and a close-up of the label on a plain background, are what we build the product reference plates from. If your photos are low resolution or inconsistent, we will tell you which extra shots would make the product more accurate.

The takeaway

Testing AI video ads pays off when you use the speed to ask better questions, not just more of them: distinct angles first, the right format for each, hooks and lengths once an angle wins, and files named so the results read cleanly. Each batch should start from what the last one proved.

If you want that loop run for your account, the AI video ads service page sets out the batch, formats and deliverables in detail, and our guide to AI UGC videos for brands covers the creator format most batches include. When you are ready, send us your current ads and results and we will come back with an angle list worth testing.