Do AI ads perform as well as real ads? How to test it

Do AI ads work as well as real ads? It depends on the ad, the offer and how you measure. Published studies point in different directions, mostly because they measure different things. This page shows what two of them found, where each is weak, and how to run a head-to-head on your own offer. The test design is the part the existing pages on testing ad hooks and making 20 ad variations do not cover: those compare ideas, and this compares how the ad was made.

The short answer: roughly equal clicks in one study, a weaker effect in another

If you want one sentence: AI ads can get clicks at a similar rate to human ads, but clicks are not sales, and the evidence on what happens after the click is thin. Do not treat either study as a forecast for your account. Use them to decide what to test.

What the studies actually measured and who paid for them

The Taboola study. Taboola, a company that places ads on news and content sites, announced its study on 28 January 2026, with researchers from Columbia, Harvard, TUM and CMU. It compared AI-made and human-made ads from the same advertiser, in the same campaign, on the same day, on Taboola's own platform. Taboola reports more than 500 million impressions and more than 3 million clicks. The raw click rate was 0.76% for AI ads and 0.65% for human ads, and the company says the two were comparable under the tightest controls. It also reports that AI ads which did not look like AI had the highest engagement, and that large, clear human faces were linked with trust.

The limits are worth knowing. Taboola sells AI ad tools, so it has a reason to like the result. A trade write-up noted that it is one platform, it matched ads rather than randomly assigning them, it published no confidence ranges, and the AI ads more often had a clear human face, which could be doing some of the work. This summary rests on the press release and a trade summary, not the full report.

The Ipsos study. Ipsos, working with Syracuse University's Newhouse School, showed 3,000 US consumers 20 ads, 10 human-made and 10 fully AI-made, from the same briefs so that the way the ad was made was the thing that changed. Human ads scored 14% higher on short-term effectiveness and 17% higher on long-term effectiveness. It measures predicted effectiveness from a survey, not real campaign results, and all the AI ads came from one set of tools. More in is AI video good enough for ads yet.

Why they disagree. One counted real clicks on a news-style ad network. The other asked people to judge ads and predicted effect. One looked at short-term action, the other also at long-term brand effect. Neither says anything certain about your offer, your audience or your price.

Why you should test your own offer

Your result depends on things no study sees: your price, your audience, how well people already know you, and whether the ad shows a real person. What AI mainly gives you is volume and lower cost per clip. So the question to answer is: "Does the AI version get a cost per lead or sale close enough to the real version that the savings are worth it?"

I would also ignore performance figures in vendor and agency blog posts that give no method. Examples are claims such as "return on ad spend within 5% across 200 campaigns" and figures quoted from Nielsen and Kantar through secondary sources only. Forum posts that quote click rates are unverified, so do not use them as data. If a number has no named study, no sample and no method, you cannot check it, and you should not plan around it.

Set up a fair head-to-head: real, AI and hybrid

Run three arms, not two. A hybrid arm often turns out to be the one worth keeping.

• Arm A, real. Footage you filmed. A phone is fine. This is your control.
• Arm B, AI. The same script made with AI video and an AI or recorded voice.
• Arm C, hybrid. A real person or real proof in the key moments, with AI b-roll and variants around them.

Keep these the same in all arms:

1. The offer, the script and the call to action.
2. The landing page.
3. The audience and the placement (for example, Instagram Reels only).
4. The daily budget per arm.
5. The start and end dates. Run all arms at the same time, not one after another.

Change only one thing: how the ad was made. Do not edit an arm while the test runs.

One confound to control: the Taboola work linked clear human faces to trust. If your real arm shows a face and your AI arm does not, you are testing faces, not production method. Put a face in all arms or in none. If a real person's face or voice is used in any arm, you need their written permission, as set out in the consent checklist.

Use the platform's own split-test tool when it has one, because it divides the audience for you so the same people are not shown two arms. Meta's rules on confidence for A/B tests change, so check the current help page in Ads Manager before you rely on any threshold it shows.

What to measure beyond clicks

Measure	What it tells you	Watch out for
Click-through rate (CTR): clicks divided by impressions	Whether the ad gets attention and a tap	Easy to win with a curious hook that brings poor leads
Hold rate (how much of the clip is watched)	Whether the middle keeps people	Short clips score higher
Cost per lead or sale	The number that pays the bills	Needs many more events to be stable
Lead quality: booked jobs, sales, no-shows	Whether the leads were real buyers	Takes days or weeks to show
Frequency and fatigue over two weeks	How fast the ad wears out	Compare arms at the same frequency
Cost per usable asset	Credits plus your hours, per ad you can run	Count rejected clips too

If a platform adds an "AI" label to one arm, note the date and watch whether that arm's CTR changes. A secondary summary claims labels lower click rates, but that is unconfirmed, so treat it as something to watch.

How much to spend, and why small tests mislead

This is the part most blog posts skip. Small tests can only catch big differences. Here is the maths so you can check it yourself. It is my own approximation using a standard two-group comparison of rates, not a number from either study.

Inputs: a baseline click rate of 0.65% for one arm and 0.76% for the other (the Taboola figures), 95% confidence (z = 1.96) and 80% power (z = 0.84). Power is the chance the test catches a difference that is really there.

Formula: the impressions needed per arm are about (1.96 + 0.84)² × [p1 × (1 − p1) + p2 × (1 − p2)] ÷ (p2 − p1)².

Working: (2.8)² = 7.84. The bracket is 0.0065 × 0.9935 + 0.0076 × 0.9924 = about 0.0140. The gap squared is 0.0011² = 0.00000121. So 7.84 × 0.0140 ÷ 0.00000121 is about 90,000 impressions per arm.

Now a made-up budget to show what that means. Say you spend $200 across three arms at an assumed $10 per thousand impressions. That buys about 20,000 impressions in total, roughly 6,700 per arm, and about 43 clicks per arm at 0.65%. With 6,700 per arm, the same formula says you could only reliably spot a gap as large as roughly 0.65% against 1.1%, a jump of about 70%. A true difference of 0.65% against 0.76% would be buried in noise. So a $200 test cannot show that AI and real ads are equal. It can only catch a large difference or a disaster.

Lead and sale counts are worse, because they are rarer than clicks. With a handful of leads per arm, one lucky day decides the winner. What to do about it:

• Look for large gaps. If AI costs more than 20% less per lead and quality holds, that is a gap worth acting on even at small scale.
• Run the full period you set, often 14 days, even if one arm leads early.
• If you cannot afford the numbers above, run the free viewer test in the AI video ad page first and treat paid results as a sanity check.

Pass and fail rules to write before you launch

Write these down before the first impression, so you do not argue yourself into a result afterwards. The numbers are examples; you set your own.

• Adopt AI for this ad type if the AI arm's cost per lead is within 20% of the real arm, lead quality is not lower after 14 days, and complaints are not higher.
• Adopt the hybrid if it beats the real arm on cost per lead, or matches it at a lower production cost.
• Keep real footage if the AI arm is worse than the real arm by more than 20% on cost per lead, or leads are worse in quality.
• Call it unclear if the gap is inside the noise. Do not claim a win. Either extend the test or move on.

When AI wins on cost per asset but not on performance, the right call may be to use it for variants and b-roll, and film the key moments. Hook ideas can be ranked with a hook rubric, and which video will perform covers five versions on one budget.

Claims you cannot check and how to treat them

• A percentage with no method. Leave it out of your plan.
• A vendor's own study. Check who ran it, how ads were matched and whether it shows uncertainty.
• "Platforms penalise AI ads." I know of no solid evidence either way. Watch your own delivery numbers.
• Fake reviews or endorsements. The US Federal Trade Commission (FTC) says endorsements must reflect the endorser's honest opinion. The FTC guidance on this does not mention AI. See fake testimonials and AI reviews, and check platform policy traps for AI ads before the ad goes live.

After the test: what to scale

If one arm clearly wins, run it longer against your next challenger. If nothing wins, make the cheaper arm your default and keep one real clip as a monthly control. Log the winner and the date. When the tool changes, the result may change too.

How I'd approach this

I would not argue about AI versus real in the abstract. I would pick one offer, make three versions, write the pass rules down, and spend enough to see a big gap, or admit the budget cannot show a small one. I would measure leads and quality, not only clicks, and I would repeat the test on a second offer before changing how the business makes ads.

Blur Studio builds the repeatable side of this for small businesses and agencies: a test log, an approval step, and a shared library of the winning clips. See the business page if that would help.