A recent Fast Company piece, Marketing attribution wasn't built for AI, makes a simple point. AI makes marketing decisions faster, but the way most teams measure results has not changed. Reports still look at channels one by one. They can give too much credit to the last click, and several platforms can claim the same sale. So AI can help a team act on an incomplete picture, only faster.
We read it as people who make ads, not as analysts. And it made us think about one question we hear more and more: how to make a 30 second AI video ad. Making it is now the easy part. Knowing whether it worked is not.
Here is a short story to show what we mean. It is invented; the brand and the person are not real. But every step in it is common.
The story: twelve ads in one week
Deniz runs marketing at a small skincare brand. Her budget is tight, and her agency fees are high, so she tries something new. With an AI video tool, a few product photos and a script, she makes a 30-second ad in two days. It looks good. She is proud of it.
Because it was so fast, she makes more. A morning version and an evening version. One with a voice-over, one with music only. A version for older customers, one for younger ones. By Friday she has twelve ads running on paid social.
To fund them, she cuts the display banners that have been running quietly for a year. They never looked strong in the reports anyway.
A month later, two things happen. Paid social reports a lot of sales. And branded search, people typing the brand name into Google, starts to fall. The search report makes search look like the problem. The social report makes social look like a hero. When she adds up the sales that each platform claims, the total is higher than the sales the shop actually made.
This is very close to the example in the Fast Company article: a budget cut in one channel shows up later as a drop somewhere else, and the channel reports point at the wrong cause. The banners she cut were probably helping people remember the brand. The new ads may have worked too. She cannot tell, because twelve versions went out at once and nothing was held back.
The problem is not the AI. The problem is that production got faster than measurement.
Why fast creative makes measurement harder
Before AI, a brand made one or two films a year. Each one was expensive, so people planned how to judge it. Now a team can make many versions in a week. Each version changes several things at once: the hook, the music, the length, the person on screen, the colour of the scene.
When results come back, nobody knows which change mattered. Was it the voice-over? The first three seconds? Or was it simply that more money went to social that month?
Channel dashboards cannot answer this alone. They were built to report on their own channel. This is why many teams are looking again at methods that measure the whole mix. Google, for example, made its open-source marketing mix model, Meridian, available to all marketers in January 2025. Tools like this are useful, but they need clean inputs: clear dates, clear spend and clear notes on which creative ran where. That last part is our side of the work.
How to make a 30 second AI video ad you can actually measure
From the production side, this is how we would plan it. None of these steps is complex; they just need to happen before the first frame is generated.
- Write one test question first. For example: does a product-first opening beat a lifestyle-first opening? One question per round.
- Build a master edit. Script, keyframes, shots, sound. This is the control version. Every variant starts from it.
- Change one thing per variant. If the test is the opening, only the first three seconds change. Same music, same length, same end card.
- Name and log every version. A simple sheet with the file name, what changed, where it ran and the dates. It sounds boring. It is the part that makes analysis possible later.
- Keep something back. Leave one region, audience or time window without the new ads, or keep the old channel running at a low level. Without a comparison, every result is a guess.
- Keep the brand look stable. Product shape, label, colour and logo should be identical in every version. If the product looks different from ad to ad, you are testing the AI, not the idea.
The craft side still matters too. AI clips need the same finishing as any film: consistent colour, a clean product, correct label text and sound that fits. We wrote about why planning still matters in why AI video ad production still needs 3D previz.
What this means for brands
Three honest points.
First, cheap variants are not free. Each extra version costs a little to make, but it also splits your budget and makes your results harder to read.
Second, do not cut a channel only because its own report looks weak. Check what happens to the rest of the mix when it goes away, ideally in a small test before a full cut.
Third, measurement starts in production. The person who makes the ad decides how testable it is. A clear master, controlled variants and a version log cost almost nothing and save a lot of argument later.
A practical takeaway
If you are about to make your first AI video ad, plan the test before the prompt. Decide what you want to learn, build one master, change one thing at a time and keep a log. Then the speed of AI works for you, not against you.
We have worked with AI imagery and video for about the last year, and with image finishing for much longer; Mehmet has 26 years in post-production. If you want to see how we approach AI video and motion work, you can find it at 35milimetre.com.