Scaling A/B Testing for YouTube Shorts

Why we built a custom app to test and learn faster

Creative Technologist Indeed 2025 Gina ContehJoseph Crist

Summary

  • Indeed's YouTube team wanted to test "brain rot"-style animations as the visuals for podcast audio clips.
  • We needed a way to test a wide range of visual variations than manual production could support.
  • Built a web app that turns automated the process of pairing podcast audio with visuals a branded YouTube Short.
  • Codifies brand guidelines directly into the tool — caption style, color palette, background animation, speaker imagery — so non-technical teammates can focus on testing variations, not production details.
  • Auto-transcribes audio via OpenAI's Whisper API and syncs a dynamic waveform visualization to it.
  • Includes a gallery of p5.js-generated background animations in brand-approved palettes, customizable speaker cards, and individual layer exports.
  • Let the team iterate and test more formats against YouTube's algorithm without losing momentum or brand consistency.

The Challenge

The YouTube team wanted to test podcast audio clips paired with 'brain rot'-style animations for YouTube Shorts. This style of animation was a first for the channel, so we needed lots of variations to test. Manually creating these assets was not in scope, so we needed to find a way to automate the process while maintaining brand consistency.

The Approach

I used Cursor to develop a simple web app that codified certain aspects of our brand guidelines. From caption style and color palette to background animations and speaker images, the web app handled the details of our brand identity allowing non-technical team members to focus on the visual varations needed to test against YouTube's algorithm.

The web app allows users to upload audio, choose a background animation, create speaker images, and generate a YouTube Shorts asset in minutes.

Features

  • Upload audio and auto-transcribe using OpenAI's Whisper API
  • Dynamic waveform visualization synced to audio
  • Gallery of generative background animations with brand-approved color palettes
  • Speaker images with customizable names and job titles
  • Individual layer exports (background, captions, speaker info, waveform)
  • Real-time preview before export

The tool created an on-brand interpretation of 'brain rot' visuals, bridging the hyperactive tone of the trend with Indeed's polished, professional identity.

I used p5.js to create the gallery of background animations.

The Impact

This simple, internal tool automated the repetitive parts of asset creation, freeing up our team to focus on the creative aspects of the project. By shifting the focus of our team from the tedious to the creative, we were able to quickly iterate without losing momentum, test more content formats faster, and maintain brand consistency at scale.