Keeping each batch of AI ads to one selling point pushed his click-through rate to 15%

Russell is a marketer who has spent the past few months making AI ads for brands. One of his clients is Knockin, which makes digital business cards and AI reception tools. By his own public account, his team can make three to five ads in one to two days, then spend a few days testing them with a small budget. In one TikTok run, a single ad reached a click-through rate (CTR) of 15.14%, with a cost per click (CPC) of about $0.06. What made the difference was writing each batch of scripts around one selling point. The AI only expands that one selling point into different ways of telling it.

The workflow at a glance

Step	What it covers	Output
Set the goal	Agree with the client on what this batch should do: get attention, build interest, build trust, or drive a sale	One-sentence goal
Set the character	Choose a main character image you approve, and write down which traits must stay fixed	Character reference and rules
Write the master script	Lock in one selling point, write one short script, and make a few key frames for the client to approve	One approved script
Expand the batch	Have the AI swap the people, scenes, and ways of telling it around the same selling point, then pick and edit by hand	Five or six short scripts
Make the videos	Break each script line into visible actions, and end with an animation that shows how the product solves the opening problem	Three to five finished videos
Small-budget test	Check the first two seconds, six-second retention, and clicks, read the comments, and decide what to change next	One test record

They make their videos with Dreamina. They use GPT and Claude to help write scripts and organize production requirements, use Claude to help build the product demo animations, and use Muse to collect real comments from the comment section.

The workflow and numbers above all come from Russell's public statements. We have not verified the ad data, and we have not run this workflow ourselves. By his own explanation, click-through rate only helps pick which versions are worth testing again. He does not yet track consistently whether they led to sales.

Why it works: three habits

Each batch sells one thing. If the first ad talks about saving time, the second about identity and the third about price, testing still won't tell you which approach works better. Keep the selling point fixed and change only the people, scenes, and openings. That way, the differences in the data actually mean something.

Write the master script first, and expand only after it is approved. One short script with a few key frames lets a client understand the idea in a few minutes. If the direction is wrong, you fix one script. You do not wait for the AI to expand six scripts, build all the frames, and then rework everything.

End by returning to the same person and problem from the start. The opening shows a nail technician who cannot reply to messages in time. The ending shows her using the business card to get the information to the customer first. By his account, one nail-tech ad performed poorly at first. It improved after he kept the original story and added a product demo animation with the same character at the end. If you swap the face or the example in the middle, viewers have to work out the story all over again.

Build it yourself: a six-step SOP

Give this to your agent. Have it check each step against the completion standard after finishing that step, then stop and wait for your confirmation before moving on.

Safety rules

• Only promise what the product can really do. The opening can exaggerate the trouble, but the product demo must match real features. Ad platforms check exaggerated claims closely.
• Get permission for any real person's likeness. Use AI-generated characters, or real people who have signed a written agreement. Do not borrow the faces of celebrities or other people.
• Set a cap on the test budget first. A small-budget test still spends real money. Set a total and a number of days for each round, and stop to review when you reach them.
• Label AI-generated content the way each platform requires.

Common pitfalls

Mixing several selling points in one batch. After the test, you cannot tell which selling point worked, so the money is spent without a clear answer. Lock one selling point per batch, and switch in the next batch.

Showing the product without showing its value. If the video shows only the interface and the logo, it proves the feature appeared, but not why it matters. Viewers need to see a specific person skip a specific hassle before they see why it matters to them.

Letting the video model generate the text. Brand names, product text, and Q&A content often come out with typos or distortion. Let the model perform the character and run it a few times, then add approved text in post-production.

Judging only by click-through rate. Good clicks with poor inquiries may point to the landing page or the product description. Look at platform data alongside feedback from clients and sales, then decide whether to change the ad, the page, or the follow-up.

Changing several things at once and blaming one of them. If the new people, scenes, and openings perform better, that only shows the whole set is worth continuing. To find out which change mattered, change one thing in the next round and test again.

To break down the structure of a viral ad first, see Hand a viral video to your agent and get it rebuilt around your product. For the product demo animation at the end, see Get Claude Code to make motion videos you can fix one line at a time.