Playbook: marketing videos, variants and honest measurement

0:00 The marketing video playbook

What if, instead of making one perfect ad, you made nine good ones in the time it used to take to make one, and let your audience tell you which works? That's the real superpower of AI in marketing video. Not one beautiful asset, but a testing system. In this lecture you'll learn which formats suit AI production, how to plan variants that actually teach you something, how to write a fifteen-second vertical ad with a shot list, and how to measure results honestly.

0:37 Why a system

Why think in systems? Because marketing is uncertain. Even experienced teams can't reliably predict which hook, voice or length will work for a new audience. Traditional production made testing expensive, so teams guessed and hoped. AI makes variants cheap, which means you can replace guessing with evidence. But cheap variants also make it easy to test sloppily, or to cross ethical lines with fake testimonials. The system keeps both speed and integrity.

1:08 The chef's test

Here's the mental model. Think of a chef testing a new dish. They don't change the recipe, the plating and the portion size all at once, because then they'd never know what diners liked. They change one thing, serve it, and watch. A variant plan works the same way. You keep a control version, and each variant changes one variable: the hook, the voice, the length, or the language. Everything else stays constant, including the real product shots and the approved offer text.

1:44 Formats that suit AI

Which formats suit AI production? Product explainers, where AI helps with voice, concept b-roll and localization, but the product interface and results come from real screen recordings. Vertical ads, where AI helps with hook and voice variants, captions and sizes, but prices and offers are real. Localized campaigns, with dubbing and native review. And founder or expert messages, where AI can help with script and audio cleanup, but the person should be filmed for real, because that format runs on trust.

2:19 UGC-style ads

And one format needs a clear warning: so-called U G C style ads, which imitate user-generated content. AI can help script and edit them. But AI-generated customers or reviews presented as real people are deceptive. The US Federal Trade Commission's rule on fake reviews and testimonials took effect in October twenty twenty-four, and the UK's advertising regulator takes fake testimonials seriously too. Use real creators, disclose paid partnerships, and label AI where required. There's no variant plan clever enough to justify a fake customer.

2:56 A variant plan

Now the variant plan itself, using a Ramadan meal-kit campaign for Saudi Arabia and the UAE on Meta and TikTok. The control: hook A, iftar in twenty minutes, a designed Arabic female voice, fifteen seconds, vertical. Variant one changes only the hook. Variant two changes only the voice. Variant three is a six-second cut. Localizations: Gulf Arabic as primary, English for expats, Urdu for South Asian audiences. Held constant: real product shots, legal-approved offer text, the call to action, caption style and the AI label policy.

3:33 Example 1: a Manchester bakery

First example, a simple one. A small bakery in Manchester makes two versions of a ten-second vertical ad for its weekend brunch. Same real footage of the pastries, same offer, same voice. Only the hook changes: fresh croissants every Saturday, versus skip the queue, preorder brunch. Each link carries a different UTM content tag. After two weekends, the preorder hook drives noticeably more orders on their site, illustratively. They didn't need a big budget. They needed one clean comparison.

4:07 Example 2: Karachi skincare

Second example, a business case with illustrative results. A Karachi skincare brand makes a master thirty-second explainer in Urdu with real product footage and a designed narrator. Then nine variants: three hooks, two voices, two lengths and an English version for the UAE diaspora. Every variant has its own UTM content value and a platform AI label where the voice is synthetic. After two weeks, one hook clearly wins on cost per purchase, while voice differences are small. They scale the winning hook, keep the cheaper voice, and record the learning in a creative log.

4:48 Watch me do it, part 1

Watch me write a fifteen-second ad as a scene table with a source column. Scene one, zero to two seconds: iftar in twenty minutes. On screen, the same words. Visual: real footage of hands opening the meal kit. Scene two, two to seven seconds: everything's measured, just cook. Visual: real ingredients laid out. Scene three, seven to eleven: fresh, local, delivered by Asr. Visual: a generated warm sunset street, for atmosphere only. Scene four: order by noon, link below, over a real plated dish and the logo.

5:26 Watch me do it, part 2

Before production, I check two things. Every claim is true and approved: the twenty minutes, the delivery by Asr, the noon cutoff. And nothing generated could be read as the product or a result. Then I set up measurement. Each variant gets a UTM content value that describes it, like hook B, voice Arabic male, fifteen seconds. I choose one success metric that matches the objective, here cost per purchase, and I agree a minimum test duration and budget before launch, so nobody declares a winner after one lucky day.

6:05 Honest measurement

Now measurement, done honestly. Match the metric to the objective: hook rate and hold for attention, click-through for traffic, and cost per key event or cost per purchase for sales. Respect consent and privacy in how you measure. For consent-aware tracking, server-side tagging and conversion APIs, the Privacy-First Measurement course goes deep, and Web Analytics with GA4 shows how to report on variants. And give tests enough time and volume. A single day's result is mostly noise.

6:38 AI and human assets together

Here's a pattern I see again and again. A London software company launches a feature with a sixty-second AI-voiced explainer using real interface recordings, a twenty-second founder clip filmed on a phone, and dubbed Arabic and Spanish versions with native review. The founder clip wins with existing customers, who value the personal touch. The explainer wins with cold audiences, who need the what and the why. AI-produced and human-filmed assets aren't rivals. They're complementary tools for different audiences.

7:12 Common mistakes

Common mistakes. Testing five things at once, so you can't tell what won. Fake customer testimonials made with avatars or generated people. Generated shots that exaggerate product results. Declaring winners after a day, or without UTMs to separate variants. And forgetting that sponsorship disclosure and AI labels are separate. Each mistake either wastes the test or damages trust, and trust is the asset you're really building.

7:41 Recap

Recap. Treat marketing video as a testing system. Keep a control and change one variable per variant. Use AI for voices, hooks, localization and atmosphere, and keep products, prices, offers and people real. Tag every variant, measure the metric that matches your objective, and give tests time. And if you want to go beyond video into interactive voice, like an agent that answers product questions or books demos, the Voice AI Agents course is your next step.

8:14 Try this now

Try this now. Take your next campaign and write a variant plan: one control and three variants, each changing one variable. Write the fifteen-second scene table with a source column, create UTM content values for every variant, and agree on the success metric and minimum test duration before you produce anything.