AI Ads: What Works, What's Legal, and How to Test Them
Most pages blur two kinds of AI ads: creative made with AI and campaigns run by AI. Here is what AI ad creative tools make well and badly, what the studies found, where the legal lines sit in the US and EU, what Coca-Cola learned, and how to test AI creative on cost per booked call.

On this page
AI ads are ads where artificial intelligence makes the creative, runs the targeting and delivery, or both. On Meta, Google, Microsoft and TikTok, machine learning already decides who sees which ad, so AI-made creative pays off mainly as a cheap way to test more angles. The angle, the offer and the proof still come from people who know the buyer, and the landing page decides whether any of it makes money.
Most pages on this topic blur those two meanings, then argue about whether AI ads "work." That question has no answer until you say which kind you mean and what you measure. If you sell a high-ticket offer, the measure is cost per booked call. Not clicks, not likes, not comments.
What are AI ads? Two different things
"AI ads" covers two layers of the same campaign. Separate them, because they fail in different ways.
| Layer | What AI does | Examples | Where it breaks |
|---|---|---|---|
| AI-made creative | Generates images, video, avatars, voiceovers and copy variants | Adobe Firefly, Runway, HeyGen, ElevenLabs, TikTok Symphony Creative Studio, the generative tools inside Meta Ads Manager | Looks synthetic, fakes proof, repeats what every competitor already says |
| AI-run delivery | Decides who sees which ad, where, at what bid and with how much budget | Meta Advantage+, Google Performance Max and AI Max for Search, Microsoft Advertising AI Max, TikTok Smart+, ads in ChatGPT | Optimizes to the wrong event, starves on thin creative, grades its own homework |
If you buy traffic on Meta or Google, you already run the second kind. Meta Advantage+ automates audience, placement and budget. Google Performance Max buys YouTube, Display, Search, Discover, Gmail and Maps inventory from a single campaign, and AI Max for Search adds broader query matching, generated ad text and automatic landing page selection. Microsoft Advertising has now opened its own AI Max to every advertiser. TikTok Smart+ automates targeting, budget, placements and creative, and lets you switch each module off, which is why we say to test Smart+ like a performance stack. OpenAI sells ads inside ChatGPT too, a channel worth piloting before you move budget.
The machine picks the audience. Your job moved upstream, into the ad itself.
Why creative became your targeting
When the platform chooses who sees the ad, the ad is the brief. Meta's Andromeda system narrows the pool of eligible ads for each person before the auction runs, and Meta built it for a world where generative AI multiplies the number of ads in the system. We broke down the mechanics in our Meta Andromeda guide. The short version: each genuinely different ad gives the system another route to a buyer.
That is the real case for AI-made creative. Not that a model writes better than your best closer. It does not. The case is that a model can turn five human angles into twenty executions in an afternoon, and AI-run delivery has more to work with when the ads say different things.
Variety means angles, not colors. Ten recolors of one hook are one ad to a buyer. These are the angles worth separating:
- Pain: what the problem costs them today.
- Outcome: the specific result they want, in their words.
- Mechanism: why your method works where the last one failed.
- Proof: a real client, a real number, a real screenshot.
- Objection: the reason they have not bought yet, such as price, time or "I tried this before."
- Identity: who the offer is for, and who it is not for.
AI can draft executions of every one of these. It cannot tell you which pains your buyers say out loud. Pull those from sales call recordings, applications and closed-lost notes, then hand the model the exact words. Then grade each AI draft against the jobs shown in the best ad copy examples.
Do AI-generated ads actually work?
On clicks, the best evidence says yes, with one condition. On booked calls and revenue, nobody has published proof either way. Here is what the two most cited studies found.
The Columbia, Harvard and TUM display ad study
Researchers from Columbia Business School, Harvard Business School and the Technical University of Munich analyzed display campaigns on Taboola covering more than 16 billion impressions and 116 million clicks from over 7,000 advertisers ("AI in Disguise" by Exner, Hartmann, Netzer and Zhang). For a fair comparison, they matched ads made by the same advertiser on the same day, in the same campaign, pointing at the same landing page, where some used AI-generated images and some did not.
- Before controls, the AI-image ads averaged a 0.76% click-through rate against 0.65% for the human-made ones.
- Within the matched tests, AI and human ads performed about the same on average.
- AI images beat human images only when they did not look AI-made. AI ads that read as AI lost the edge.
- Large faces and polished aesthetics read as human to viewers. Intense color saturation read as AI.
- In purchase campaigns, the AI images showed no significant effect on conversions, and the authors note their data did not cover landing pages, so conversion effects could not be cleanly pinned on the images.
One caution before this goes into a pitch deck: the data comes from native display placements on Taboola, not Meta or TikTok feeds.
The MIT IDE personalized video experiment
Madhav Kumar, a fellow at the MIT Initiative on the Digital Economy, and Anuj Kapoor of the University of Missouri ran a WhatsApp field experiment with more than 21,000 consumers of an Indian retailer of eco-friendly products (MIT IDE summary). Personalized videos with an AI avatar beat personalized image ads on click-through rate by 9.4 percentage points and generic videos by 6.5 points. The authors flag two limits. It was a single exposure, so novelty may explain part of the lift. And the marketing team kept full control of the message, using AI as a production tool rather than the creative engine.
What the research means for a high-ticket funnel
Both studies stop at the click. Your business does not. A click that books nothing is a cost, so read the research as permission to test, not proof of profit:
- AI creative can match human creative on attention when it does not look synthetic.
- Faces carry trust. Use real ones for anything that implies a real person or a real result.
- The page after the click still decides the economics.
That last point is where most of the money hides. Dillon Kivo sells a high-ticket offer through a book-a-call funnel fed by Meta ads. With the funnel we built and the ads we ran, his booked calls went from 44 to 94, comparing two one-month windows, while unique visitors rose only about 10%, because landing page conversion went from 5.3% to 10.3%. The traffic barely moved. The page did the work. If your ads already get clicks, fix the landing page your paid traffic hits before you fund more creative.
AI ad creative tools: what they make well and what they do not
AI ad creative tools generate the ad itself: images, video, avatars and copy, usually from your website, a product photo or a script. "Ad creative AI" is also a brand search, because AdCreative.ai is a tool with that name. These tools are good at volume, sizes and fast variations of an angle you already have. They are weak at what sells a high-ticket offer: a sharp angle, believable proof and a face the buyer trusts.
Four kinds cover most of the market.
| Kind | Examples | Makes well | Falls short on |
|---|---|---|---|
| Static ad generators | AdCreative.ai | Image ads in every size from your site and brand kit, copy variants, competitor ad research | Template layouts, generic copy, scores that are predictions rather than results |
| Generative AI inside the ad platform | Meta Advantage+ creative: text and image generation, background generation, image and video expansion, image animation | Variations of your own asset, fitted to each placement, inside the system that buys the media | Some enhancements are on by default, and the preview does not show every version |
| Video and avatar tools | Creatify, Arcads | Link-to-video, UGC-style avatar ads, translations, many variations at once | AI actors that read as customers, disclosure duties, testimonial rules |
| General image models | Adobe Firefly and the partner models inside it | Scenes, backgrounds, b-roll stills, motion statics | Invented details, licensing that depends on the model |
Static ad generators
AdCreative.ai is the clearest example. It scans your web page and brand kit, then produces image ads for each platform and format, ad copy, product photoshoots and short videos, and it scores creatives before launch (AdCreative.ai). Appier announced in February 2025 that it was acquiring the company (Appier's announcement). The homepage claims over 90% accuracy for its performance prediction and does not explain on the page how that is measured. Treat any pre-launch score as a guess about attention, not a forecast of booked calls.
Good for: resizing a winning ad for every placement, quick headline variations, product-led offers. Weak for: a $10K offer sold on trust. Templates produce template ads, and your buyer has scrolled past that layout before.
Meta's generative AI features in Ads Manager
The AI creative tool most advertisers already have is Meta's own. Advantage+ creative uses AI to generate and enhance variations of image, video and carousel ads, and Meta's page names text generation, image generation and touch-ups, background generation for Advantage+ catalog ads, image and video expansion, and image animation (Meta Advantage+ creative). Meta also says some enhancements may be on by default, that you can switch them off, and that the previews are samples rather than every variation that can run.
That last point is the catch for an expert brand. A generated line or background can run under your name without anyone on your team reading it. We recommend keeping the format fixes, such as expansion to fit a placement, and reviewing generated text and images like any other copy before launch. Ads made with Meta's generative tools can also carry an "AI info" label, covered in the platform rules below.
Video and avatar tools like Creatify and Arcads
Creatify turns a product link into a video ad, makes avatar videos, batch-creates a dozen variations at once and shows spend, ROAS and CTR for each creative (Creatify). Arcads sells UGC-style video built on a library of more than 1,000 AI actors, custom avatars that can hold your product, and translation into more than 30 languages (Arcads). Its terms make the customer responsible for any duty to disclose AI-generated content when posting to TikTok or X, and bar using its videos for misleading or deceptive advertising (Arcads terms).
Good for: testing hooks and scripts before you book a shoot, product and software demos, localizing a founder's message. Weak for: anything that looks like a customer talking. An AI actor saying "this program changed my business" presents an experience nobody had. The FTC does not ban AI avatars, but a testimonial is illegal when the experience behind it is fake, avatar or not (see the FTC rules below). Let the avatar speak as the brand, never as a client.
General image models
Adobe Firefly creates and edits images, video, audio and vector graphics, and Adobe says its own Firefly models are trained on licensed Adobe Stock content and public domain content whose copyright has expired (Adobe Firefly). The same app also runs partner models from Google, OpenAI, Runway and others. Adobe's training-data statement covers its own models, so check which model made an image before it goes into an ad.
Good for: backgrounds, scenes, b-roll stills and motion statics behind a real voice or a real headline. Weak for: anything that stands in for evidence. Image models invent details, such as a dashboard that does not exist or a number on a screen nobody approved, and in a proof-led ad an invented detail is an invented claim.
Where AI creative falls short
- The generic look. In the Columbia, Harvard and TUM study, AI images beat human ones only when they did not look AI-made (the research above). Template layouts and too-perfect scenes give it away fast.
- Invented claims. Generators fill gaps with numbers, guarantees and outcomes you never gave them. Every claim in an AI-made ad needs evidence behind it, the same as a human-written one.
- Fake people. AI avatars posing as clients, AI testimonials and synthetic before-and-after shots carry the legal risk, and they spend the trust an expert brand sells on.
- Disclosure. Political and social issue ads, EU audiences and some platform tools carry disclosure rules, laid out under is AI advertising legal.
How to test AI creative against human creative
Run the tool's output through the cost-per-booked-call test below, with three additions for tool-made ads:
- Same angle, two makers. Give the tool an angle you wrote from sales calls, and run its version next to a human-made version of the same angle. Tool against tool tells you nothing about your buyer.
- Ignore the tool's score. A predicted performance score and the platform's click-through rate are hints. The verdict is cost per booked call, then show rate and close rate.
- Log what was switched on. Record which Advantage+ creative enhancements ran on each ad, so you know whether you tested your creative or Meta's variation of it.
Count the whole cost too: the subscription and the hours spent fixing outputs belong in the cost of the creative. When the AI version wins, keep the angle and reshoot it with real proof before you scale it.
Is AI advertising legal?
Using AI to make an ad is not illegal in itself in the US or the EU. The rules attach to deception, disclosure and political content. This is a practical summary, not legal advice, so run anything borderline past counsel.
United States: the FTC
The FTC's Rule on the Use of Consumer Reviews and Testimonials took effect on October 21, 2024 (FTC questions and answers). It bans fake reviews and testimonials, including AI-generated ones, attributed to people who do not exist or never used the product. The FTC's guidance adds the part most ad buyers miss: the rule has no blanket ban on AI avatars. An avatar delivering a testimonial breaks the rule only if the underlying testimonial is fake or false, though an avatar can still be deceptive under the FTC Act. Using a celebrity avatar to praise a product without permission breaks the rule when people would believe the celebrity endorsed it.
Coaches, consultants and course sellers should also know Operation AI Comply, the FTC sweep against business-opportunity schemes that promised AI-powered income from online stores. The agency's line was blunt: there is no AI exemption from the laws on the books. If your ad says AI will make the buyer money, treat it like any income claim. If you cannot back it up, it does not run.
Platform rules: Meta, Google and TikTok
- Meta requires a disclosure when a social issue, electoral or political ad uses a photorealistic image or video, or realistic audio, that was digitally created or altered to show a real person saying or doing something they did not, a realistic person or event that does not exist, or altered footage of a real event. Undisclosed ads get rejected, and repeat failures bring penalties.
- Meta also labels ads created or materially edited with its own generative AI features with "AI info." When the tool generates a photorealistic person, the label sits next to "Sponsored."
- Google requires election advertisers to tick the "Altered or synthetic content" box when an ad depicts realistic synthetic people or events, and generates the disclosure automatically on some formats.
- TikTok labels content from its Symphony Creative Studio as "AI-generated" when its ad policy requires disclosure.
European Union: the AI Act
The transparency rules in Article 50 of the EU AI Act apply from August 2, 2026 (European Commission timeline). Two parts touch advertisers:
- If you publish a deepfake, meaning AI-made or altered image, audio or video that resembles real people, places or events and would pass as authentic, you must disclose that it is artificial. Evidently creative, satirical or fictional work carries a lighter duty.
- Companies that build generative tools must mark the output in a machine-readable way. Certain tools already on the market before August 2, 2026 have until December 2, 2026 to comply.
The rules worth running every account by:
- If a testimonial did not come from a real customer with real experience, it does not run. AI or not.
- If an ad shows a real person saying something they did not say, it does not run.
- If an avatar speaks, it speaks as the brand, never as a client.
- If the ad touches politics or a social issue, use the platform's disclosure every time.
- If you sell into the EU, label anything that would pass as real footage.
Is Coca-Cola using AI ads?
Yes. Coca-Cola remade its 1995 "Holidays Are Coming" commercial with generative AI in 2024, then released new AI-made versions in 2025 with the AI studios Secret Level and Silverside AI. Both rounds drew backlash. Critics called the 2024 film "soulless," and after the 2025 launch, the media intelligence firm CARMA measured positive conversation falling from 23.8% to 10.2%, while negative conversation held near a third, 31.4% before and 32% after (Marketing-Interactive).
Coca-Cola's own position is that human creativity stays at the heart of its advertising while AI handles production. That split is right. The lesson is about where the AI showed. A soda brand can absorb a rough week of sentiment. A consultant or expert brand sells trust, and synthetic-looking creative spends exactly that. Use AI where the buyer does not need to believe a human was there. Keep it away from anything that stands in for proof.
How to test AI ads: the workflow we judge on booked calls
AI ad testing means two things. It is testing ads made with AI against ads made by people, and it is using AI tools, including the ad platform's own, to build variants, split budget and read results. For a high-ticket offer both meanings end in the same place: the verdict comes from booked calls and closes, not from a tool's score or the click-through rate.
Here is the workflow we run when AI creative goes into a high-ticket account. It works on Meta, TikTok and Google video, and it keeps the platform from grading its own homework.
- Write one hypothesis per test. Name the angle, the one thing that changes and the result you expect, in cost per booked call. "An AI-made motion static of the mechanism angle books calls within 20% of the founder-to-camera version" is a hypothesis. "Let's see if AI ads work" is not. Pull four to six distinct angles from sales calls, applications and closed-lost notes, each naming a different pain, outcome, mechanism, proof point or objection.
- Build the variants: AI makes executions, people write the angle. Let AI generate what is cheap to vary: hook lines drawn from your own call transcripts, scenes, backgrounds, b-roll, motion statics, every ratio and placement size, dubbed versions of real founder video. Keep human: the angle, the offer, every claim and number, testimonials and any face that implies a real client. Produce two or three executions per angle: one human-shot (founder to camera or a real client), one AI-assisted, and optionally a hybrid. Every angle gets a human control.
- Split the budget so the test is fair. Change one variable and freeze everything else: same offer, same landing page, same form, same calendar for the whole window. Change the page mid-test and you learn nothing. For a clean answer on an expensive decision, use the platform's split test, which keeps the audience groups apart (see the tools below). For a quicker read, give one angle's executions their own test ad set with a capped share of spend, so your proven ads keep running. Let each angle spend about three times your target cost per booked call before you judge it. Less than that is noise.
- Send the right signal. Fire the standard Lead event only on a submitted application, and pass booked and showed calls back to the platform. Optimize on a cheap action and the algorithm finds you cheap people.
- Run it long enough for calls to happen. Our default is at least seven days, so every weekday is in the data, and then until the spend rule above is met. Calls land days after the click, so follow the leads each ad produced in the window until their calls are held. A call booked for next Tuesday is not a no-show today.
- Read results on booked calls and closes, not clicks. Judge on cost per booked call first, then show rate, then close rate and cash per call. Use click-through rate only to diagnose a hook. If the application answers and the call outcome do not sit on the same CRM record as the ad that started it, fix that before you test anything.
- Kill fast, scale the angle. Kill an angle that stays above your control's cost per booked call after its fair budget, and kill any ad whose bookings do not show or never close, however cheap they look. When an AI execution wins, promote the angle, not the asset: reshoot it with real proof, move it into the main campaign and write the next hypothesis from what it taught you.
| Metric | What it tells you | Decision rule |
|---|---|---|
| Click-through rate | Whether the hook stops the scroll | Never crown a winner on CTR alone. Modest CTR with cheap booked calls is a keeper. |
| Landing page conversion rate | Whether the ad's promise matches the page | If it drops for one angle only, fix the message match, not the ad. |
| Cost per booked call | What a real sales conversation costs | The test metric. Kill angles that stay above your control after a fair budget. |
| Show rate | Whether the ad pulled buyers or browsers | Cheap bookings with a poor show rate mean the angle sells curiosity, not intent. |
| Close rate and cash per call | Lead quality | The final word. A cheap call that never closes is the most expensive one you bought. |
Show rate is a funnel problem as much as an ad problem. Confirmation pages, reminders and pre-call content carry most of it, which our booking funnel guide walks through step by step. When Meta is the channel, our Meta ads management for high-ticket offers is built around this loop.
AI ad testing tools and workflows
The most useful AI ad testing tools are the ones you already have, inside the ad platforms. Three on Meta matter for this workflow.
- A/B test in Ads Manager. Meta's A/B testing page describes splitting your audience into random groups that do not overlap, changing one variable between otherwise identical versions, and comparing up to five ad variants. Use it when the decision is expensive, such as AI-made against human-shot creative for your main angle.
- Split tests through the API. Meta's split testing guide for the Marketing API says tests divide audiences into mutually exclusive groups, supports creative tests of two to seven cells with one ad each, and, for a conversion test, treats the version with the lowest cost per action as the best performer. It also notes that larger reach, longer schedules and bigger budgets give more statistically significant results. That verdict is only as good as the event you optimize for, which is one more reason Lead must mean a submitted application.
- Advantage+ creative. It generates and adjusts variations of your ads automatically, as covered above. That makes it a production tool, not a test: log which enhancements ran, or you cannot tell whether your creative or Meta's version of it won.
Third-party AI ad testing tools mostly do one of four jobs. Pre-launch scoring predicts attention or performance from the creative before it runs. Synthetic audience tools ask simulated personas to react to ads in place of a survey panel. Creative analytics tools pull ad-level data from your accounts and tag each creative by hook, format and angle, so patterns show up across tests. Automated rules move budget or pause ads when a metric crosses a line you set. Each can save hours. None of them knows which calls closed unless you feed it your CRM data, so treat its output as the source of your next hypothesis and let booked calls and closes decide.
What percent of ads are AI now?
No reliable figure exists for the share of all ads made with AI. Platforms do not publish one, and surveys measure intent, not output. The closest verified data points:
- In the IAB's 2025 Digital Video Ad Spend and Strategy report, 86% of video ad buyers said they use or plan to use generative AI to build video creative, and buyers projected that gen AI creative would reach 40% of all ads by 2026. That is a forecast from buyers, not a count.
- In the Columbia-led study above, by the end of a 73-day window that spanned the July 2023 launch of Taboola's free generative ad tool, 6.25% of advertisers had used it at least once, and it produced 3.04% of ad images after launch.
Treat any "percent of ads" figure quoted without a source as marketing. The number that matters is whether your AI-assisted ads win on booked calls.
Where AI belongs in your ads, and where it does not
| Use AI for | Keep human |
|---|---|
| More executions of a proven angle: formats, ratios, hooks, captions | Choosing the angle and the offer |
| Backgrounds, b-roll, product scenes and motion statics | Testimonials, client results and screenshots |
| Dubbing and localizing real founder video | Any face that implies a real client or a real result |
| Script drafts built from your own call transcripts | Final copy, claims and compliance sign-off |
| Resizing and versioning for every placement | Income and outcome promises |
That split is how we approach ad creative production: people own the angle and the proof, AI scales the executions, and the numbers pick the winners. When the founder is the face of the offer, real video still does work no generator can, which is why VSLs and testimonial shoots stay in the plan.
Getting clicks but not booked calls? Run a free landing page audit and check whether the page is the leak before you blame the creative.
Frequently asked questions
What is AI ad testing?
AI ad testing means two things: testing ads made with AI against ads made by people, and using AI tools to run the tests, from generating variants to splitting budget and reading results. Meta's own A/B test splits your audience into groups that do not overlap and compares up to five versions of an ad. For a high-ticket offer, judge every test on cost per booked call, show rate and close rate, not on clicks or a tool's predicted score.
Is AI advertising illegal?
No. Using AI to make an ad is not illegal in itself in the US or the EU. What breaks the law is deception: the FTC's rule on reviews and testimonials, in force since October 21, 2024, bans fake testimonials including AI-generated ones, and the EU AI Act requires deepfakes to be disclosed from August 2, 2026. Meta also requires an AI disclosure on political and social issue ads, and Google does on election ads. This is general information, not legal advice.
Do AI-generated ads actually work?
On clicks, yes, when they do not look AI-made. A Columbia, Harvard and TUM study of real display campaigns found ads with AI-generated images performed about the same as human-made ads on average, and beat them only when the images did not read as AI. No published study shows AI creative improves booked calls or sales, so test it against a human-made control and judge it on cost per booked call.
How long should an AI ad test run?
Our default is at least seven days, so every weekday is in the data, and until each angle has spent about three times your target cost per booked call. Then follow the leads from that window until their calls are held, because bookings land days after the click. Ending a test on day two usually crowns the ad with the cheapest clicks, which is rarely the ad that books buyers.
What is the best AI for ad creative?
There is no single best one, because the tools do different jobs. Static generators such as AdCreative.ai turn your website and brand kit into image ads, Meta's Advantage+ creative generates variations inside Ads Manager, video and avatar tools such as Creatify and Arcads make UGC-style video, and general image models such as Adobe Firefly make scenes and backgrounds. The best one for your account is the one whose ads win on cost per booked call against a human-made control, so test two and let the ad account decide.
Is there a free AI commercial generator?
Yes. TikTok's Symphony Creative Studio is free for TikTok advertisers and turns a product URL or script into a video, with digital avatars and dubbing. Meta Ads Manager and Google Ads have generative tools built in, and standalone tools such as Runway and ElevenLabs offer free plans with limited usage. Free gets you executions, not a winning angle.
Can I use an AI avatar or an AI testimonial in my ads?
An AI avatar, yes. A fake testimonial, no. The FTC says its review and testimonial rule has no blanket ban on AI avatars, but an avatar testimonial is illegal if the underlying testimonial is fake or false, and an avatar can still be deceptive under the FTC Act. Let the avatar speak as your brand, never as a customer who does not exist.
Written by
Mohamed Ali Naaoui
Funnelslayer
We build conversion systems for coaches, consultants and brands already paying for traffic: offer, copy, design, build, tracking and the optimization after launch.





