How to make UGC ads with AI

Know the rules before you start

The US Federal Trade Commission's rule on fake reviews and testimonials, in effect since October 21, 2024, bans testimonials that misrepresent who is giving them or whether that person exists, and it applies to AI-generated content. An AI presenter saying "I have used this for three months and my skin cleared up" is a fake testimonial. An AI presenter demonstrating the product and stating facts you can back up is an ad.

So write scripts as demonstration and information, not as a personal customer story. TikTok, YouTube and Meta each have rules for labeling realistic AI-generated content, and in the EU the AI Act's transparency duties for synthetic media apply from August 2, 2026. If you base the presenter on a real person, get their written consent. This is general guidance, not legal advice.

Steps 1 and 2: build the presenter and the product

1. Create one presenter and keep them. Pick a look that fits your buyer: age range, style, setting. Build a reference set of three to five images in a casual outfit: front, three-quarter and profile, lit like a phone selfie near a window. A recurring presenter builds familiarity the way a real creator does.

2. Build the product reference set separately: hero angle, front, back and a sharp logo close-up, plus one image of the product in a hand for scale. Small text on the label remains the weak point of current models, so plan shots where fine print is not readable.

Step 3: script in three parts

3. Write the ad as three separate clips, which is how you will generate and test it.

Hook, about 3 seconds. One line and one visual that stops the scroll. For example: "Three things this bottle does that yours does not," while the presenter holds the bottle close to the lens.

Demo, 5 to 10 seconds. Show the product doing the thing: pouring, opening, applying, folding. One or two actions per clip. Every spoken claim must be true and provable.

Call to action, 3 to 5 seconds. What to do next and why now. Product held next to the face, label to camera.

These lengths fit current models. Kling 3.0 Omni generates up to 15 seconds per request, according to Kling's guide as of October 2026.

Step 4: generate with UGC settings

4. Set the aspect ratio to 9:16. Write prompts that read like a phone video: "Handheld selfie angle, slight movement, she sits on a sofa in a bright living room and turns toward the lens. Daylight from a window on the right." Avoid cinematic moves; dolly shots and orbits look like a commercial and lose the UGC feel.

Plan who appears in each clip. If one shot needs both the presenter and the product, you need a model that takes two subjects in one request, such as Kling's Omni models with two elements. Otherwise, alternate: presenter shots from the presenter set, close product shots from the product set, with a hand described in the prompt.

Keep the setting consistent within one ad. A cut from a living room to a bathroom mid-sentence breaks the illusion of a single recording.

Step 5: test hooks, keep the middle

5. The hook decides much of an ad's performance, so make several. Generate five to ten hooks that vary the opening line, the first visual and the setting, and attach each to the same demo and call to action. Run them as separate ads with the same budget and compare hold rate and click-through.

Once a hook wins, vary the demo: a different room, another use, a second presenter. Because the presenter and product come from fixed reference sets, the variations look like the same creator and the same product, which keeps the test fair.

Add voice and captions in editing. Use the same voice across variants, and burn in captions, since many viewers watch muted.

Checklist and making it in RefVid

Checklist: no fake testimonial claims; consent for any real likeness; presenter and product sets built; hook, demo and CTA scripted; 9:16; handheld, natural-light prompts; five or more hooks; captions burned in; AI label where required.

In RefVid, save the presenter as a character and the product as a product, each with 1 to 5 images. Each video uses one asset, generated on Kling 3.0 Omni in 9:16, 16:9 or 1:1. Ten 3-second hooks cost 1,260 credits at 42 credits per second. RefVid clips are silent, so voice, music and captions happen in your editor.