AI Ad Creative: How We'd Build an Ad Worth Testing

If you're running Meta ads and every new version feels like the last one, we'd start with the idea. What does someone need to understand before they'd consider buying? Let's pick one question and build the ad around it. Then we can try a different opening without starting the whole thing over.

Say you've already changed the actor, rewritten the hook, and tried a different background. Those changes might be worth testing, but the ad could still be making exactly the same argument. So before we ask AI for another batch, let's figure out whether we need a new message or a better way to explain the one we've got.

We'll use a made-up label printer to work through that decision, then look at how the same approach fits a software demo. These are worked examples and our recommendations for making creative. We haven't run a campaign experiment for this guide, so we can't tell you which version will win.

Are we changing the idea, or just the opening?

Imagine we're making an ad for a small label printer that connects to a phone. Someone looking at it might be wondering how hard it is to set up. Someone else might want to know what comes in the box. And another person might just want to see what the labels look like on their pantry jars.

We'd make different ads for those questions. The setup ad needs to show the app and the printer working together. The unboxing ad needs to show what's included. The pantry ad needs a close look at the finished labels on real containers.

That's what we mean by a concept: the main idea the ad helps someone understand. A variation keeps that idea and changes part of how we deliver it.

So if we keep the setup demo but open on the finished label instead of the phone, we've made a variation. If we switch to showing everything in the box, we've moved to a different concept. Both can be useful. We'd just give them different names so we know what we're trying to learn.

If you're still deciding how to present the idea, our UGC ad examples guide walks through formats such as demos, tutorials, and unboxings.

Before we make more ads, what's actually going wrong?

If an ad has had a few rough days, we'd look at the rest of the buying journey before blaming the creative. Did the offer change? Does the landing page still load and show the right product? Is tracking working? Those are things to check with whoever runs the account.

You can keep making creative while you investigate. The point is to give the next ad a specific job. If buyers keep asking whether label rolls come with the printer, an unboxing could answer that. If the problem is a broken checkout, another opening line won't repair it.

And if the idea still makes sense, we wouldn't throw it away just because we've looked at it too many times ourselves. We might try a clearer first shot or a simpler explanation. That's a much more focused brief than “make something fresh.”

Let's turn that question into a brief

For our example, let's assume the product page confirms four things: the printer works with a Bluetooth phone app, supports labels up to 12 mm wide, charges over USB-C, and includes one starter roll. Extra rolls cost more, and there's no desktop-printing feature.

Now suppose we're choosing the setup angle. Here's the brief we'd give the person or tool making our ad:

Who we're talking to:
Someone who wants labels for small containers but isn't sure how the printer works.

Question the ad should answer:
How do I make my first label from my phone?

What we'll show:
The real app opening, a label being designed, the printer producing it,
and the finished label on a container.

Facts we can use:
Bluetooth phone app, labels up to 12 mm wide, USB-C charging,
and one included starter roll.

Details to keep accurate:
The app screens, printer, label size, included items, and current offer.

What we won't claim:
Desktop printing, free extra rolls, or a setup time we haven't measured.

Next step:
View the printer and current offer on its product page.

The useful part is “what we'll show.” It gives us something to build around and something to check when the video comes back. “Make setup look effortless” leaves too much room to skip a step or invent a feature.

If your starting point is a product page, we've put a longer version of this brief in our product URL to video ad guide.

Where we'd use AI, and where we'd bring the real footage

AI ad creative can include copy, images, voices, presenters, or video scenes. You don't have to use it for every part of the ad.

For this printer, we'd bring real footage of the app and the printing step. Then we could use AI to help write the opening, make a surrounding scene, or add narration. If we're asking someone to judge how the printer works, we'd want them to see that printer doing the work.

The same goes for software. Let's say you're advertising a scheduling app. You want to show how someone sets office hours, creates a booking link, and shares it. We'd record those steps in a clean demo account first. Then we'd build the introduction, voiceover, captions, and ending around that recording.

A generated dashboard might look tidy, but you'd have to check every button and result against your actual app. Starting with the screen recording gives us the right thing to explain. We can trim the pauses and zoom into the important parts while keeping the steps intact.

We'd also keep the presenter honest. They can explain what the product does without pretending they've personally used it for months. The FTC's guidance on reviews and testimonials distinguishes the use of an AI avatar from the truth of the testimonial underneath it: avatars aren't automatically prohibited, but fake experiences and deceptive portrayals can still create problems. We checked that guidance on October 4, 2026.

So, what should we change in the next version?

First, get one version into a state you'd actually be comfortable running. Watch it through, check the product details, and make sure the script says what you meant. This becomes the version we compare the others with, often called the control.

For our printer demo, we'd start with two possible openings:

• Open on the phone: Show the label being designed, then follow it through to printing.
• Open on the result: Show the finished pantry label, then go back to how we made it.

We can keep the rest of the demo, offer, and next step the same. Now there's a clear reason for each version to exist: we're exploring whether it helps to see the process or the result first.

You could try a different narrator next. Or shorten a part that drags. We'd make those separate changes so reviewing the videos stays manageable and we can describe what changed.

That doesn't turn a live campaign into a clean experiment by itself. Different audiences, placements, delivery, and landing pages can affect the result. Keep a note of those conditions when comparing performance, and don't assume the opening caused every difference you see.

How we'd put this together in Cospark

We build Cospark, so this is our own product workflow. The brief and examples above can also be used with another tool or a human editor.

If you've already written the script, start with Script to Video. If you're starting from one product page, URL to Ad gives you that starting point. Either way, bring the buyer question, the facts you can support, and the images or footage the ad needs.

We'd work on the first version before requesting a whole batch. Read the script and look through the scene plan. If the brief says the printer includes one roll and the script says “everything you'll ever need,” fix that line before making more videos around it.

Then watch the result. If one scene is wrong, ask for that scene to be changed and check that the parts you liked survived the edit. Once the first ad makes sense, give the next version a specific instruction, such as “open on the finished label and keep the rest of the demo the same.”

For a closer look at preparing product images, see our product-photo video workflow.

Give the finished ad a proper watch

It's easy to notice a strange face or an awkward cut. We'd also pause on the less obvious things: the label on the product, the price in the caption, and the screen that's only visible for a second.

Here's the short check we'd use before handing it over:

• Does the product or interface match what someone will actually buy?
• Can we support every feature, offer, and experience mentioned in the script?
• Are the captions, product details, and next step readable on a phone?
• Do we have permission to use the footage, music, voice, and likenesses?
• Have we checked the platform's current disclosure requirements and the landing page?

Google also tells advertisers to review its generated images for accuracy and misleading content before publishing. Its generated-image guidance, checked October 4, 2026, is a useful reminder that a finished generation still needs a person to look at it. For platform-specific preparation, see our Facebook AI-ad workflow and TikTok AI-ad workflow, and check the current platform rules before launch.

If something fails that check, fix it before multiplying it across more versions. We'd rather review one corrected scene now than discover the same wrong product detail in a whole batch.

What we'd keep track of

When you look back at the work, count the ads that are ready to use. Keep the rejected generations and editing time in the record too, because they're part of what it took to get there.

You don't need a complicated spreadsheet to start. Note which buyer question each ad answers, what you had to fix, and what you spent on generation and editing. If you're recording your own time, keep the hours separate from cash costs unless you've chosen an hourly rate.

That gives you a useful production record. Campaign results are a separate question: an ad can be accurate and well made without persuading the audience you put it in front of.

For your next ad, we'd start by writing down the one thing a buyer needs help understanding. Find something real you can show them, then build the explanation around it. Once that version works as an explanation, try a different opening. You'll have a much clearer brief for AI, and a much clearer reason for making the next video.