What the score is
The Opportunity Score orders emerging topics by how attractive they look to a creator deciding what to make next. It runs from 0 to 100.
What it is not: a prediction of views, a guarantee, or a calibrated probability. It is ordinal, not cardinal — a 90 is more attractive than a 70, but it is not “1.28× better”.
The formula
Five components are multiplied together, then adjusted by how much we trust our own measurement:
- Momentum — how fast the topic is growing, and whether it is still accelerating. A large but slowing topic ranks below a smaller climbing one.
- Outlier strength — the median amount by which the topic’s videos beat their own channels’ normal performance, compared at matched video age.
- Demand signal — evidence that an audience wants this, measured per video rather than in total. Three videos pulling 40k each is a stronger signal than ten splitting 50k.
- Early-mover advantage — decays with topic age, because the value of this product is arriving early.
- Competition adjustment — penalises topics already covered, weighted by the size of the channels covering them.
Why multiplied rather than added
These are gates, not contributions. A weighted sum would let a dead topic score respectably because competition happened to be low — which would be actively misleading advice. Multiplying means any component near zero collapses the score, which is the correct behaviour.
Why confidence sits outside
Confidence does not describe the opportunity. It describes how much we trust our measurement of it. A topic built on three videos and one hour of observation may genuinely be excellent, but we cannot yet tell — so the score is shrunk toward the middle rather than allowed to post a 97 on noise.
Outlier detection
A video is not impressive because it has many views. We compare each video against the median performance of its own channel’s recent uploads at the same age — never against lifetime averages, which would make every new video look like a failure.
Where a channel has too few comparable uploads, we report insufficient baseline rather than inventing a multiple, and exclude that video from topic-level medians. Counting it as average would drag every topic toward 1.0 and suppress exactly the breakouts on smaller channels that matter most.
Known limitations
- The weights are informed priors, not values fitted to measured creator outcomes. They will be tuned as real performance data accumulates.
- “First detected” is when we saw a topic, not when it started. A topic may circulate in communities before any video appears.
- Competition is measured in YouTube supply only. Search difficulty and recommendation dynamics are not modelled.
- Opportunity windows are estimates derived from observed decay, presented as ranges. They are not deadlines.
- Clustering is imperfect. A badly merged topic produces a confidently wrong score, which is why cluster confidence feeds the confidence adjustment.
Every score stores its components, so any number in the product can be broken down into the signals that produced it.