Case Studies

YouTube Virality Analysis

Turning 300,000+ YouTube videos into a measurable model of what actually drives views — built for 1of10.

OVERVIEW

What Drives YouTube Performance at Scale

Built for 1of10, a platform that helps creators understand why some videos explode and others stall. The goal was not a deck of interesting findings. It was a system that produces measurable business outcomes.

We assembled a dataset of more than 300,000 YouTube videos spanning many channels, niches, and audience sizes, then used BigQuery, Vertex AI, and Python to turn unstructured metadata, titles, and thumbnails into measurable signals creators could act on.

The scale was necessary, not academic. Virality is a rare event, and you need enough signal to separate the real drivers from noise.

The Challenge

Most content-strategy advice is subjective and anecdotal. The work was to replace opinion with measurement:

  • Collect and process video-level metadata at scale

  • Engineer features that actually predict performance

  • Normalize results across channels of very different sizes

  • Find measurable indicators of virality

  • Turn raw YouTube data into decisions a creator can make

The pipeline had to handle large-scale ingestion while staying flexible enough for fast experimentation.

Links

APPROACH

The Approach

I built an analytics and ML workflow that:

  • Ingested YouTube metadata across 300,000+ videos

  • Processed titles and thumbnail features, including OCR and vision signals

  • Generated structured performance metrics normalized per channel

  • Created derived engagement and virality indicators

  • Modeled the relationship between metadata and view performance

It combined statistical analysis, feature engineering, and ML experimentation to surface the patterns with real signal rather than coincidence.

Core architecture:

  • Python data pipelines for ingestion and processing

  • BigQuery for analytical storage and large-scale querying

  • Feature-engineering workflows producing 300+ signals

  • Thumbnail analysis using Google Vision and OCR

  • Vertex AI for ML experimentation

  • Automated reporting and analysis layers

The design left room for the obvious next step: turning the analysis into recommendation and optimization models.

Technical Stack

  • Python

  • BigQuery

  • Vertex AI

  • Google Vision

  • SQL

  • Data Visualization

  • OpenAI API

  • Google Cloud Platform

RESULTS

Key Outcomes

  • Processed 300,000+ YouTube videos

  • Engineered 300+ metadata features

  • Identified measurable patterns tied to viral performance

  • Informed strategy that contributed to 200,000+ additional views

  • Helped drive acquisition of 100+ new customers

Why This Matters

Media companies increasingly compete on data-driven content optimization. This project showed how analytics and AI can:

  • Reduce creative guesswork

  • Sharpen content strategy with evidence

  • Scale experimentation past gut feel

  • Build a repeatable growth framework

The same approach transfers to marketing, e-commerce, and any creator business where the inputs are unstructured and the outcome is a rare event.

Looking Forward

Planned next steps:

  • AI-generated title recommendations

  • Thumbnail scoring models

  • Real-time performance forecasting

  • Recommendation systems for creators

  • Multi-platform virality benchmarking

Interested in building something similar?

We help organizations design scalable AI and analytics infrastructure for:

  • forecasting

  • growth analytics

  • operational automation

  • machine learning platforms

  • agentic AI workflows

  • cloud-native data systems

Let’s build systems that turn data into operational leverage.