Subscribe
AI for App GrowthFacts as of

AI App Ads: We Tested 20. 0 Were Usable

The short answer

Can AI make app ads you can actually run? A 20-ad test

Not as they come out. In our test, AI tools made 20 app ads in about 4 minutes for under $3, and none was ready to run: they invented reviews and ratings, faked before and after photos, and made up logos. Letting AI make only the scene, then adding the app's real screen, icon and our own words, made 4 of 4 usable.

Key facts

Chapters

  1. 0:00Can AI make app ads you can run?
  2. 0:44How did we test AI ads?
  3. 2:29How fast can AI make app ads?
  4. 3:22Do AI ads make things up?
  5. 4:50Are AI before-and-after ads honest?
  6. 6:00Can AI keep your brand straight?
  7. 7:37Are AI ads worth it at all?
  8. 8:47How do you make AI ads you can run?
  9. 10:12The AI ad checklist

Transcript

Edited for reading

Teal is a fictional AI photo editor. Every ad shown is real AI output from our own test, labelled "AI-generated example": GPT Image 2.5 (OpenAI) and Higgsfield Marketing Studio for 16 images, Kling 3.0 for 4 five-second videos, one prompt each, run on 3 October 2026. Appelune makes its videos with Higgsfield; no tool maker paid for or reviewed this test.

Can AI make app ads you can run?Watch from 0:00

We asked AI to make twenty ads for an app. It finished in about four minutes. They cost under $3. And at a glance, they look great.

Then we checked every one against the rules an ad has to pass before it runs. None of them was ready to run.

We asked for a customer review. It made one up, signed by a person who doesn't exist, without a word of warning. Four ads showed a before and after that wasn't even the same photo.

So what's actually wrong with AI-made app ads? And is there a way to use them that works? By the end, you'll have the method that took us from zero usable to four out of four.

Our test app is Teal, an AI photo editor we made up. The ads are real AI output, and every number has a source.

How did we test AI ads?Watch from 0:44

First, why bother. Ads for apps eat creative fast. In AppsFlyer's 2025 report, big non-gaming apps averaged about 2,400 video ad variations a quarter. Singular puts the median advertiser at 13 new creatives a week.

Few small teams can keep up with that by hand. That's the promise of AI. So we tested it.

Here's the brief. Teal fixes blurry, dark photos in one tap, with a seven-day free trial.

We wrote twenty ad ideas a small app team might try. A before and after. A question. A meme. A free-trial offer. A creator holding a phone.

Three were the risky kind real teams still brief: a review quote, a star rating and a comparison with other apps. We asked for them plainly, the way a busy marketer would.

The tools were GPT Image 2.5 from OpenAI, and Higgsfield's Marketing Studio, for sixteen images. Then Kling 3.0 for four five-second videos. Appelune makes its videos with Higgsfield. Nobody paid for this test.

One prompt per ad. No retries. A one-hour clock.

Every ad then faced seven checks. Correct text. The topic clear in a second. Honest claims. The right brand. The right shape for its placement. A clean image. A believable app screen. The videos got one more, for motion.

Two AI graders checked every ad. One ran the test. The other never saw the prompts. Where they disagreed, the stricter grade stood.

We judged the first impression at feed size, about as big as an ad shows on a phone. Then we zoomed in for the details.

How fast can AI make app ads?Watch from 2:29

All sixteen images came back in under two minutes. The four videos took about two more. Twenty ads, about four minutes.

And honestly, our first reaction was, wow. The lighting is good. The type is clean. The phones look real. One would pass for an agency's work.

Ad two is a hand holding a phone with a blurry birthday cake on it. Ad six is a party meme that's actually funny. Ad thirteen puts a phone on a sunny desk beside a coffee. If you scrolled past any of these in a feed, you wouldn't stop and think, that's AI.

54 credits for all twenty. At Higgsfield's listed prices, that's about $2 to $3. A creator video on a UGC platform like Trend costs about $70 to $90 each.

Our take: speed and price aren't the problem anymore. They haven't been for a while.

But a good-looking ad isn't the same as one you can run. So we started checking.

Do AI ads make things up?Watch from 3:22

The first trap is the worst one: invented proof.

Ad three is a review card. Five stars, a smiling headshot, and a quote: Teal saved my wedding photos, signed Sarah M. Sarah doesn't exist. Neither does the review.

Ad seven gives Teal 4.9 stars from 50,000 reviews. It calls Teal the number one photo fixer on the App Store. Teal has no rating at all.

Then the subtle one. A creator in her living room, holding up her phone, with a caption: this app fixed all my blurry photos. She's AI too. A made-up person telling a made-up story about your app.

Yes, we asked for these. But the tools made every one without a word of warning. A busy marketer could ship them by lunch.

And they don't look like mistakes. They look like the review cards and rating badges you see in real ads. That's what makes them dangerous.

The platforms have rules for exactly this. Meta's ad standards ban exaggerated claims about results, and products with features that don't exist. And Apple's rule 2.3.1 covers marketing outside the App Store too. Misleading ads can get your app removed and your developer account closed.

Meta labels ads made with its own AI tools that show AI-made people. And it says it'll start detecting ads made with outside AI tools.

Our take: an AI tool will write any claim you ask for. Checking it is still your job.

But the second trap hit a photo app harder.

Are AI before-and-after ads honest?Watch from 4:50

For a photo app, the before and after is the whole pitch. Our prompts asked for the same photo, before and after.

Look closely at ad one. Same group of friends, same skyline. But the people have moved. Heads tilt differently. Teal can't re-pose people. That's not a fix. It's a different photo.

Ad four hides it better. Its phone shows a lake split down the middle. But the mountains, the shoreline and the trees change across the line.

Ad eleven adds confetti to the after side that isn't in the before at all. Ad twelve gives the woman a different face. And when we turned ad twelve into a video, a grey band flickered down both photos.

Four images, plus that video, failed the same way. The AI drew an after that looks better than Teal's fix could.

In ad sixteen, and the video made from it, the blurry puppy is already perfectly sharp. There's nothing to fix.

That's the trap. Show a result your app can't deliver, and you're advertising a feature that doesn't exist.

Our take: if your ad shows a result, use your app's real output. Never a picture of what you hope it does.

Then there was the problem nobody asked for.

Can AI keep your brand straight?Watch from 6:00

We gave the tools one brand colour and one name. No logo and no screens, the way most people prompt.

So eleven of the sixteen images invented a logo for Teal. At least seven different ones. A leaf, a magic wand, a lotus, a star. None of them is Teal's icon.

We counted an invented logo as a fail, because a made-up logo isn't your brand. People learn your icon from ads, then look for it in the store.

It also drew app screens that don't exist. Buttons and sliders Teal doesn't have.

Then the small stuff. One headline printed twice. A garbled letter in the name. Camera labels that turn to gibberish when you zoom in.

The videos had their own problems. Two kept the squarer shape of their start image, not the tall story shape we asked for. Two morphed. A phone screen warps at three seconds. A puppy's toy changes shape.

Even our own grading had a lesson in it. The grader that ran the test passed one ad, the low light story. The blind grader spotted confetti that only exists on the after side. So the stricter grade stood. If one careful reviewer can miss it, a busy team will too.

So here's the score. Twenty ads. How many could you run as they came out? Zero.

Four were fixable in a few minutes: swap in the real logo, fix a letter. Sixteen were unusable. Without the logo rule, one more, the blurry cake, would have passed.

Our take: the tools aren't bad at pictures. They're bad at the parts that are yours. Your claims, your screens, your logo.

Which raises the obvious question. Are AI ads worth it at all?

Are AI ads worth it at all?Watch from 7:37

The early evidence says they can work. But it's mixed.

Taboola and a group of university researchers studied native ads at scale. They compared AI-made ads with human-made ones from the same advertisers. Clicks came out about the same.

But in an Ipsos survey of brand ads, the human-made ones scored 14% stronger in the short term, and 17% stronger in the long term.

Neither one tested app install ads. So read them as a direction, not a verdict.

Meta said in April that more than 8 million advertisers had used one of its AI creative tools. It says advertisers using its AI video tool saw over 3% higher conversion rates. That's from its own tests.

And volume alone isn't magic. Singular found creative volume on its own doesn't move results, unless spend scales with it. In AppsFlyer's data, the top 2% of non-gaming creatives took 43% of the spend.

Think of it as a numbers game with a quality floor. More shots only help if every one of them is safe to run.

Our take: AI's real gift is more good shots at a winner. Not more ads.

So we spent the rest of the hour on that. Could we get ads we'd actually run?

How do you make AI ads you can run?Watch from 8:47

Round two used one simple idea. Let AI make the part it's good at, the scene. Keep the parts that make claims for yourself.

So the new prompts said no text, no logos, and a flat green screen on the phone. Four scenes: a night party, a cosy sofa, a cafe table and a concert. Four credits.

Then we put Teal's real screen into each phone. Its real icon. And our own words, typed, not drawn by AI.

The party ad says night photos, fixed in one tap. The sofa ad asks about blurry pet photos. The cafe ad shows a pet photo ready to fix. The concert ad promises dark shots, fixed.

The blind grader passed one of the four straight away and flagged three. Their screens showed the step after the fix, and one "fixed" photo still looked blurry. Fair point. We swapped in Teal's Fix this photo screen. Then all four passed.

Notice what changed. Nothing the AI made can make a claim now. Every word, every screen and every logo is ours. The AI only made the room around the phone.

Four of four, in about twenty minutes from the first prompt, for about a credit an ad.

The catch is honest too. Our compositing was code. In a design tool, it's a few minutes of a person's time per ad.

Our take: AI is a fast photographer. It's a terrible copywriter for your claims.

The AI ad checklistWatch from 10:12

One: ask AI for scenes only. No text, no logos, no app screens.

Two: put your real screenshots in the phone, and your real icon on the ad. That's the brand check, solved before it starts.

Three: write every word yourself. Every claim must be true today, with no ratings, reviews or rankings you can't show.

Four: show results only from your app's real output, never an AI after. Run a real photo through your app and use that.

Five: before anything runs, check the seven points. Text, topic, honesty, brand, shape, a clean image, a believable screen.

Then make the variations from there. New scenes, new hooks, new words, around the same honest core. That's how a small team gets near those 13 new creatives a week.

Do that, and AI earns its place. Lots of real shots at a winner, for pennies and a few minutes each.

What changed in AI for app growth lately, and which of it matters for your app? That's ChatGPT Intelligent UI: Which Apps Lose?.

Sources

  1. AppsFlyer, 2025 Creative Report (high-spending non-gaming apps, 2,365 video variations a quarter; top 2% of creatives took 43% of spend) (29 April 2025)
  2. Singular, Creative Benchmark Report 2026 (median 13 new creatives a week; volume alone doesn't move results)
  3. Meta, Q1 2026 earnings call (8 million+ advertisers using its gen AI creative tools; 3%+ higher conversion in its video generation tests) (29 April 2026)
  4. Trend.io pricing (UGC videos about $69 to $92 each), checked 3 Oct 2026
  5. Higgsfield pricing (about 3 to 5 cents a credit), checked 3 Oct 2026
  6. Meta Transparency Center, Unacceptable business practices
  7. Apple, App Store Review Guidelines 2.3.1(a)
  8. Meta Newsroom, GenAI transparency for ads (updated 1 Jun 2026)
  9. Taboola with Columbia, Harvard, TUM and CMU researchers, AI vs human ads (native ads, 500M+ impressions) (28 January 2026)
  10. Ipsos with Syracuse Newhouse, AI Ads Are Good Enough, And That's the Problem (20 ads, 3,000 US adults) (18 May 2026)

Caveat: Neither the Taboola nor the Ipsos study tested app install ads. Our test is one app, one run, first takes only; different prompts or tools will give different results. Round one counts an invented logo as a fail, a rule we set after seeing the ads; without it, 1 of 20 passes.

Watch next