How AI short video lifted new-product CTR from 1.2% to 4.8%
Same product, same price, only the creative and the receiving page changed — and landing-page CTR moved from 1.2% to 4.8% in four weeks. The four actions that did the work, and what the 35 discarded assets revealed.
First, where the numbers come from: 1.2% was the click-through rate on a new product landing page before optimisation, and 4.8% is the same metric for the same product four weeks after the videos went live. Only the creative and the receiving page changed — the product, the price and the placements did not — which is what makes the result worth examining.
Four actions produced most of the gain. In order of effect: the first three seconds, real usage context, distributing tailored versions per platform, and cutting half the assets based on data.
Three metrics that get confused
| Metric | What it measures | Optimising it means |
|---|---|---|
| Click-through rate | Share of viewers who click | Whether the hook works |
| Completion rate | Share who watch to the end | Whether the first 3 seconds match the length |
| Conversion rate | Share of clicks that convert | Whether the landing page matches the creative |
A CTR move from 1.2% to 4.8% does not mean orders quadrupled. If the landing page cannot hold the traffic, the gain becomes wasted clicks. The video and the landing page have to change together.
1. Script structure: the first three seconds decide the rest
The structure is simpler than people expect; the difficulty is execution:
- 0–3s — ask a specific, uncomfortable question. Not "do you want more sales" but "can your customer see the fabric on a phone in three seconds?"
- 3–8s — give the counter-intuitive conclusion. Break the assumption that everything must be photographed, and state the cost difference
- 8–20s — show the process, not the result. The generation process carries information and reads as more credible than a finished image
- Final 3s — one call to action. One link, never three options
The most common mistake is spending the opening three seconds on a logo build and music. That is the most expensive real estate in the video, spent on self-introduction.
2. Generation: keep humans at the judgement step
Split the work deliberately, or the output looks synthetic:
| Step | Owner | Why |
|---|---|---|
| Product selection and selling points | Human | Requires commercial judgement |
| Script structure | Human | This is what drives CTR |
| Visual generation | AI | Repetitive work, cheaper and faster |
| Model and style swaps | AI | Removes the cost of a shoot |
| Multi-language versions | AI | Many languages in one pass |
| Final review and compliance | Human | AI cannot judge exaggerated claims |
The principle: AI provides volume, people provide judgement. Let an AI decide which selling point to lead with and the output homogenises quickly.
3. Distribution: do not post one cut everywhere
The same video performs very differently across platforms, usually because of length and opening tempo. We ended up with a per-platform version set:
- Vertical short-form: 15–25s, hook inside 2 seconds
- Horizontal platforms: 30–45s, keeping more process detail
- Visual communities: video stills plus long copy to catch search traffic
- Own site: embedded in the product page as supporting detail, not judged on completion
Producing three lengths from one asset costs little and avoids forcing one cut across every channel.
4. Review: cut half the assets
Over four weeks we produced 46 assets and kept 11. Two criteria decided it:
- Three-second retention above that account's baseline
- Post-click dwell time meeting the bar, which shows the landing page holds
Failures were stopped outright, not optimised again. Volume is not the asset; the handful that reveal the pattern is.
Five common mistakes
- Opening with a logo or a music bed
- Using an AI-generated model to show detail that must look reliable
- Three different calls to action in one video
- Optimising the video while ignoring landing-page consistency
- Watching total views instead of three-second retention
FAQ
Does AI-generated video look fake?
It depends what you use it for. It works well for scenarios, backgrounds and multi-language versions. It exposes itself when you use it to show detail that must be credible, such as stitching or fabric texture. We keep real footage for critical detail and use generated assets for context.
Which categories suit this approach?
Those where the visual difference is easy to show and the unit price does not justify a real shoot. Apparel, accessories, home goods and small appliances fit. Categories whose value lies in complex function need real footage more.
How many assets before a pattern appears?
Start with at least 20. Below that, single-asset noise hides the signal. Of our 46, the first 15 were close to worthless and the real pattern only emerged after number 20.
How long does a CTR gain last?
Creative fatigues, typically within four to eight weeks. Sustaining it means continuous iteration, which is exactly why standardising the production process beats chasing one viral asset.
To see how the AI tools fit into that pipeline, look at AI e-commerce tools; for a content diagnosis of your own category, contact us.