Methodology

Every historical video is expanded into derived metrics: views per subscriber, like rate, comment rate, combined engagement rate, title length in characters and words, and duration in minutes.

Raw view counts are misleading on their own, because a channel with two million subscribers will out-view a channel with twenty thousand regardless of quality. The core normaliser is therefore views per subscriber, taken on a log10 scale so that the distribution is roughly symmetric.

Each video receives an audience-adjusted performance percentile: 75% from the log views-per-subscriber rank and 25% from the engagement-rate rank.

When you simulate a video, the engine compares your inputs against every historical video and computes a similarity score from subscriber scale (log distance), category, video length and chapter usage.

The most similar videos are kept as a weighted neighbourhood. Their audience-adjusted performance is averaged using similarity as the weight, so close matches count more than distant ones.

Alongside the neighbourhood estimate, the input is scored on independent factors: title strength, thumbnail quality, length fit for the category, topic demand within the dataset, audience scale and structural signals such as chapters.

Title strength rewards a workable character count, concrete numbers, curiosity or emotional framing, and penalises all-caps shouting and vague phrasing. Thumbnail quality is measured in the browser from the uploaded image: contrast, brightness balance, colour saturation, visual complexity and subject prominence.

The final score blends the comparable-video estimate with the factor scores, then is clamped to 0–100.

Confidence reflects how much evidence supports the estimate: how many close matches exist, how tightly they agree, and whether a thumbnail was supplied. Few matches or widely scattered outcomes lower confidence even when the score is high.

A high score with low confidence means the inputs look promising but the dataset has little precedent for this combination.

Each recommendation targets the weakest factors and carries an estimated point impact derived from re-running the scoring model with that single change applied. The Improvement Studio applies selected recommendations together and re-scores, so the projected score is computed rather than assumed.

This is a correlational model built on a modest dataset. It cannot observe the two variables YouTube weighs most heavily (click-through rate and average view duration), nor can it see the recommendation system, upload timing relative to a trend, or audience loyalty.

Patterns such as 'videos with chapters show higher engagement' are associations within this sample, not proof of cause. Treat the score as a structured second opinion on the controllable inputs, not a forecast of view count.

Accuracy improves as the dataset grows and as the videos in it resemble the ones you are planning.