Do AI video ads work?
Last updated 2026-10-20
The short answer
Sometimes, and it depends on what a study measured. In click tests, AI-made ads have matched human ads or beaten them. In the one video test we found, human ads scored higher on Ipsos' effectiveness indexes. In another study, telling people an ad was AI-made cut clicks. AI helps most with speed and testing. Trust comes from a real product and real people.
What the studies found
NYU Stern and Emory, November 2025 (visual ads). Researchers Ghose, Lee, Todri and Adamopoulos compared ads made by human experts, human ads modified with generative AI, and fully AI-generated ads. Fully AI-generated ads raised click-through rate by up to 19%. Using AI to modify human ads showed no significant gain. Telling consumers an ad was AI-made cut click-through by 31.5%. The Stern summary gives no sample size, and it describes visual ads without saying whether they were images or video.
Taboola with university researchers, January 2026 (ads on Taboola's platform). The study covered more than 500 million impressions and 3 million clicks, comparing matched AI and human ads from the same advertiser, campaign and day. Under the tightest statistical controls the two performed comparably. The raw averages were 0.76% for AI ads and 0.65% for human ads. AI ads that did not look AI-made had the highest engagement. This is Taboola's own press release, and it does not give the number of ad pairs.
Ipsos and Syracuse University, May 2026 (video ads). 3,000 US consumers saw 20 ads: 10 human-made and 10 AI-generated from briefs reverse-engineered from the originals, using AI tools with no human creative intervention. Human ads scored 14% higher on Ipsos' Creative Effect Index (short-term sales impact) and 17% higher on its Equity Effect Index (long-term brand health). Only 25% of viewers of AI ads were at least somewhat confident the spot was AI-made. The strongest AI ads were direct and product-driven, and the authors say the gap may narrow.
NIQ, December 2024 (AI-generated ads of mixed quality). More than 2,000 participants watched AI-generated ads, and about 150 of them had their brain activity measured with EEG. NIQ reported that viewers found the ads less engaging, and that even ads rated high quality produced weaker memory activation than traditional ads. NIQ's release does not describe the ads or the method in detail.
Meta, September 2024 (its generative AI ad features). Meta says that on average, campaigns using its generative AI ad features had an 11% higher click-through rate and a 7.6% higher conversion rate than campaigns that did not. Meta published no method, sample or time period, so read it as a company average, not a test of AI against human creative.
Make ad variants from your own footage in Woven Video
Woven Video is an AI video editor for Mac. The research above favors a simple use of it: start from footage you shot, make several versions to test, and generate a shot only when you are missing one.
- Re-cut the opening. Drag a clip's edge to trim the first shots, split at the playhead with Command B, or ask in chat for a version that opens on a different line. Captions are made from your speech, and Woven keeps transcripts by file content, so captioning the same audio again after a re-cut costs nothing for those seconds.
- Size it for the placement. In the Format menu pick Vertical 1080x1920, Square 1080x1080 or Horizontal 1920x1080, or Custom for ratios such as 4:5. Export is an MP4 at 1080p or 1440p from the Export button, and 1080p when the export is requested in chat.
- Keep a real person on screen. Put a facecam over your product shots: a smaller video on a higher lane, with a Circle mask and a border, so the person stays visible while the product plays.
- Fill gaps by generating. Ask for an image, a video clip or music. Generation uses Woven credits, and Woven has no stock footage or music library, so a missing shot is either yours or generated.
- Study an ad you admire. Give Woven's agent the link to a competitor's ad and it saves the video, studies its frames and spoken words, and reports back on the hook, format, pacing, shots and sound. Only the opening 90 seconds of a long video get a breakdown, and a site that needs a login or blocks your region may refuse the download.
Still your job. Woven does not publish to ad platforms. Export the MP4s and run the test, and set any AI label, in the platform.
How to build a set of variants worth testing
- 1Start from real footage. Film the product and the kind of person who would use it. Taboola's study found a large, clear human face among the strongest trust signals, and Ipsos' strongest AI ads were direct and product-driven. Use generated shots to fill gaps, and label them as AI where the platform asks.
- 2Spend the opening on the point. Meta's ad guidance says a video should catch attention within 3 seconds and carry its key message or brand. TikTok says to state the core proposition in the first 3 seconds and land the hook within 6. In Woven, trim the opening or ask in chat for a cut that starts on your strongest line.
- 3Change one big thing per variant. TikTok advises 3 to 5 creatives per ad group and variations that differ meaningfully. Change the first shot, the opening line or the length, not a caption color. Re-captioning a re-cut costs nothing for audio Woven has already transcribed.
- 4Export each one and test it in the platform. Name each MP4 for what changed, upload them, and see which holds attention. When delivery trends down, add fresh creatives. That decline is creative fatigue.
Woven runs on macOS 15 or later and needs a Woven plan or trial. Captions and generation use Woven credits.
FAQ
The studies point to habits, not a formula. Show the real product: in the Ipsos test the strongest AI ads were product-driven and direct. Show real people, or label the AI: in Taboola's study AI ads that did not look AI-made had the highest engagement, and a large, clear human face was among the strongest trust factors. Win the opening: Meta and TikTok both point to the first 3 seconds. Then test variants against each other.
On production, the evidence we found is positive. In a field experiment by Ju and Aral on arXiv, 2,234 participants worked in human-human or human-AI teams and produced 11,024 ads. Human-AI teams made 50% more ads per worker. Their text was rated higher, but human-only teams' images were rated higher, and the AI-assisted ads were more similar to each other. In the study's test on X, better text raised click-through and better images lowered cost per click, so check both in your own account.
The studies disagree on whether people notice, and agree the response is cooler. In Ipsos' test only 25% of viewers of AI ads were at least somewhat confident the spot was AI-made, yet when viewers compared the two directly they rated AI ads as less creative, less emotionally engaging and less informative. In NIQ's 2024 study, participants picked out most of the AI-generated ads, rated them less engaging, and even the high-quality ones produced weaker memory activation. The two used different ads and methods, so do not read them as one number.
Follow the platform. In the NYU Stern and Emory study, telling consumers an ad was AI-made cut click-through by 31.5%. That is a cost worth knowing, not a reason to skip a required label: TikTok says it rejects or restricts ads with undisclosed AI-generated content. The practical answer is to build on real footage and real faces, so fewer shots need a label. The legal page covers the rules.
TikTok's creative guidance, updated June 2025, suggests 3 to 5 creatives per ad group and making them meaningfully different when testing. It says to refresh when delivery shows a steady decline, and to add new creatives to the existing ad group.
More answers
Is it legal to make AI ads, and can AI edit my video? See Is it legal to make AI ads?, Can AI edit my video? and Can you monetize AI videos on YouTube?.
Sources
- AI Ads Are Good Enough, And That's the Problem (Ipsos, May 2026), checked October 20, 2026
- The AI Advertising Paradox (NYU Stern Research Highlights), checked October 20, 2026
- New Study: AI Ads Match Human Creative (Taboola press release, January 2026), checked October 20, 2026
- NIQ Research Uncovers Hidden Consumer Attitudes Toward AI-Generated Ads (NIQ, December 2024), checked October 20, 2026
- Meta's AI Products Just Got Smarter and More Useful (Meta, September 2024), checked October 20, 2026
- Collaborating with AI Agents: Field Experiments on Teamwork, Productivity, and Performance (Ju and Aral, arXiv), checked October 20, 2026
- Creative best practices for performance ads (TikTok for Business), checked October 20, 2026
- Tailored Campaigns (Meta for Business), checked October 20, 2026
Try Woven free for 3 days
Then $25/mo billed yearly ($299/yr) or $39/month. Native Mac app for short-form video.
Paste into Claude Code to install and connect Woven.