AI Product Video Ads: A UGC-Style Workflow for Ecommerce Sellers

You can make a decent product video ad with AI help in an afternoon, as long as you use AI for the script, the shot list and the edit, and keep your real product on camera. Ask a video model to invent the product and you get a lookalike with drifting logos and a sleeve that changes length mid-clip. This post is a workflow for a small store: write the script from facts, film a few real shots on your phone, and use AI where it is actually reliable.

Status of everything below: the prompts are untested templates. I did not generate or edit any video for this post, so there are no result clips and no performance claims. The one-line examples are illustrative, written by hand.

What does "UGC-style" mean, and what should stay real?

UGC-style means the ad looks like something a customer filmed: vertical, handheld, natural light, someone talking to the camera, no studio polish. The style works because it feels like a person, which is exactly why it is easy to abuse.

Draw a line before you start:

If you want to understand how video models behave and which clips to avoid prompting, see the video generation prompts guide for Sora, Runway and Kling. For hook and caption wording, the Reels script prompts post covers a lot of ground, and the ad creative prompts post covers offers and angles. This post is the production workflow that ties them together for a product page and a feed ad.

Step 1: build a one-page fact sheet and claims list

Every good script starts from facts you can prove. Reuse the fact-sheet idea from the Shopify product description prompts, then add two lists the video needs.

FACT SHEET
Product: [NAME]
What it is in one line: [TEXT]
Materials / specs I can prove: [LIST]
Price and any offer: [Rs. X, offer terms and end date]
Delivery and return terms: [YOUR ACTUAL POLICY]
Who it is for / not for: [TEXT]

DEMONSTRABLE ON CAMERA (things a phone can show in one take)
- [e.g. pours without dripping]
- [e.g. fits a 15 inch laptop]

CLAIMS I MUST NOT MAKE
- No health, safety or "results" claims
- No numbers I cannot prove
- No "customers love it" unless I can show real reviews

The "demonstrable on camera" list is the secret of ads that work without special effects: one visible proof per shot. If you cannot film it, it does not go in the script.

Step 2: prompt for hooks and a 20 second script

Give the model the fact sheet and constrain the structure. Ask for options, because the first hook is rarely the best.

You write short vertical video scripts for a small ecommerce store.
Use ONLY the facts in the FACT SHEET. If a claim is not there, do not write it.
Never write a first-person claim of using or owning the product (no "I've been using this for months"). The speaker is the store owner or a presenter introducing the product, speaking as the seller.

[PASTE FACT SHEET]

Produce:
1. 8 hook options (first 2 seconds, max 12 words each), covering: a problem, a number I can prove, a demo, a comparison to the usual option, a question.
2. For the 3 strongest hooks, a full 20 second script in this table: second range | what the viewer sees | spoken line (max 14 words) | on-screen text (max 6 words).
3. A 7 second cut of each script that keeps only hook, one proof, and offer.
4. A list of every factual claim in the scripts, each tagged with the fact sheet line that supports it.

Item 4 is your audit trail. If a claim has no supporting line, delete it. This is the same discipline as in the hallucinations lesson: the model fills gaps with plausible things.

An illustrative script row (hand-written, not model output):

0-2s | Hand pours tea from the bottle into a cup, no drips | "Pour without the drip." | POUR. NO DRIP.
2-6s | Close-up of the lid and spout | "Spout and lid tested on my own counter." | (none)

Notice the second line is a weak claim if you did not actually test it. Edit lines like that out or turn them into something you did do.

Step 3: shoot the real footage with a shot list

Ask the model for a shot list built from the script, in the order that is easiest to film, not the order of the final edit.

Turn this script into a phone shot list for one person filming alone.
Group shots by location and setup so I film each setup once. For each shot: shot number, setup, framing (vertical 9:16), how long to hold, what I say if anything, and the script second it maps to.
Add 2 spare cutaway shots for every 20 seconds of script (hands, texture, packaging).
Flag any shot that needs a second person, a tripod, or special lighting.
SCRIPT: [PASTE]

Practical filming rules that no prompt can replace:

• Film near a window, not under a ceiling light.
• Hold each shot a second longer than you think you need. Cutting is easy, extending is not.
• Record clean audio separately for the voiceover, even a quiet room and your phone's voice memo app.
• Shoot the product label and the price offer in focus as their own shots.
• Film the 7 second version's shots first, then the extras.

Step 4: where AI genuinely helps in the edit

Use AI for tasks where mistakes are obvious and fixable:

• Captions. Generate subtitles from your voiceover and fix every proper noun and number by hand. Wrong captions on a price are a real cost.
• Cut suggestions. Paste your transcript with timestamps and ask which sentences repeat, which are weak, and where to trim for a 7 second cut.
• Text overlays. Ask for three on-screen text options per shot, each under six words.
• Voiceover. A synthetic voice reading your approved script is common. If you use one, think about whether a viewer would assume it is a real person's testimony, and label it where the rules require.
• B-roll that is not the product. A clearly decorative background loop is lower risk than a generated product shot, but see the labelling section below.

A prompt for the edit stage:

Here is a transcript with timestamps of my 22 second video ad.
1. List sentences that repeat an idea or add no proof.
2. Propose a 7 second cut: which timestamps to keep, in order.
3. Write 3 on-screen text options for each kept segment (max 6 words, no claims beyond what I said).
4. List every number or product name in the transcript so I can check the captions.
TRANSCRIPT: [PASTE]

Where I'd skip AI video generation: the product itself, anyone's hands holding the product, text on packaging, and anything that demonstrates function. The failure checklist in the AI product photos workflow (logos redrawn, proportions shifting, parts added or removed) applies to video too, and it is worse because artefacts move.

Format and platform specs: what I verified

Spec sheets for Meta are inconsistent, and I could not open Meta's own spec pages from my environment. A search summary of third-party guides on 2026-10-09 agreed on a 9:16 vertical canvas, 1080 by 1920 as a common size, and disagreed on safe-zone margins (values like roughly the top 14 percent and bottom 35 percent appeared in some, other guides gave different numbers). Treat none of that as verified. Do this instead:

1. Upload a test clip to your ad account and use the in-product preview for each placement.
2. Keep captions and your price in the middle of the frame, away from the top and bottom edges where the app puts its own buttons.
3. Check Meta's current Ads Guide in your ad account for the file size and length limits of your placement.

For your own store, I could read Shopify's product media page (2026-10-09). It says uploaded videos can be up to 10 minutes, up to 1 GB and up to 4K (4096 x 2160), in .mp4, .mov or .webm, that Shopify converts and serves them as .mp4 or HLS, and that you can embed YouTube or Vimeo instead. The add-media page adds that a product can hold at most 250 images, 3D models or videos in total, that you cannot upload more than 1,000 videos in a week, and that videos per store and video storage depend on your plan (the Starter plan listed 250 videos and 50 GB at the time I read it). Confirm the numbers for your plan in your admin, because plan tables change. Your theme must support video; the add-media page links to a list of supported themes and says older theme versions may need updating.

A cheap practical tip follows from this: compress the file before upload even though the cap is generous. A 40 MB vertical clip loads far faster on a mobile connection than a 400 MB export, and the product page is where shoppers decide.

Labels, disclosure and testimonial rules

Three separate things get mixed up, so keep them apart.

1. Is it a fake testimonial? The US FTC's endorsement guidance Q&A (read 2026-10-09; it applies to US law, so use it as a principle, not Indian law) says an endorsement must reflect the honest opinion of the endorser, that you cannot talk about your experience with a product you have not tried, and that material connections such as payment should be disclosed clearly and conspicuously. An AI persona saying "I switched to this and my skin cleared" fails the first test in any country.

2. Does the ad need an AI label? On Meta, a search summary of Meta's help page and its February 2025 announcement on 2026-10-09 says Meta labels ads made or significantly edited with its own generative AI features, shows "AI info" in the "About this ad" menu (sometimes next to "Sponsored"), and has said it will detect third-party AI content through industry-standard provenance signals. Minor edits such as resizing or colour correction are not labelled. The rollout timing of third-party detection was unclear in what I could read, so check Meta's help page on AI images in ads yourself. In India, a search summary of coverage on 2026-10-09 says ASCI issued its guidelines on labelling synthetically generated content in advertising on 29 September 2026, to take effect three months after publication (one trade outlet gives 1 January 2027, so confirm the date on ASCI's own text). It describes a risk-based approach: fabricated endorsements and unauthorised deepfakes are prohibited even with a label, labelling is mandatory where AI materially influences purchase decisions (for example synthetic visuals used to show product performance), and minor enhancements need no label. The summary lists example labels such as "Audio/Video created using AI". I read ASCI's homepage and influencer page, but not the guidelines themselves, so read the PDF and do not rely on my paraphrase.

3. Is a creator or a friend paid or gifted? ASCI's influencer page (read 2026-10-09) says it is important for brands and influencers to disclose material connections, and points to guidelines introduced in May 2021 at asci.social for the details. If you work with real creators, the safest default is a clear, upfront disclosure in the video, not buried in a caption.

None of this is legal advice. If your product category is health, cosmetics, finance or children's goods, the rules are tighter, so ask a professional.

A real-creator alternative that still uses AI

Many small stores do better with a cheap, honest pipeline: send the product to five real customers who agreed to be filmed, and use AI only to prepare. Have the model draft a one-page brief with talking points as questions ("what did you use it for first?") rather than scripted claims. Ask for their permission in writing, keep proof of any gift or payment, and disclose it in the ad. The review mining prompts can help you find the questions customers actually ask, which makes the best talking points.

Test the ad without fooling yourself

Run a simple comparison in your own ad account. Make two or three versions that differ in one thing (the hook, or the length), keep the offer and the audience the same, give them equal budget, and wait for enough orders or clicks to see a clear difference. I won't quote a typical lift or watch time because I have no data for your store and found none I trust. If a result is within the noise, call it a draw and test something bigger, such as the offer.

A short pre-publish checklist:

1. Every claim maps to a line on the fact sheet.
2. The product on screen is the real product, not a generated lookalike.
3. No synthetic person claims personal use.
4. Price, offer end date and delivery terms are on screen long enough to read, and match the product page.
5. Captions checked for numbers and brand names.
6. Labels or disclosures added where the current platform and ASCI rules call for them.
7. Product page has the same video or a clear photo of the item, so the ad and the page agree.

For the product page itself, the Shopify product description prompts pair well with the video. If your sales also run through marketplaces, the Amazon, Flipkart and Meesho listing prompts cover those channels, whose rules differ from your own store's.