Views and Downloads Move Together — But Neither Measures What You Think
Two counters, one moderate correlation, and a median that makes the whole discussion look different.
| Metric | Value |
|---|---|
| Correlation (views ↔ downloads) | 0.46 |
| Median views per skill | 2 |
| Median downloads per skill | 5 |
| Skills with zero views | 46 (7.8%) |
| Skills with zero downloads | 8 (1.4%) |
The medians are the story
A median of two views per skill is the number that should reset your intuitions about this directory. Half the skills have been looked at twice in their lifetime.
Meanwhile the median skill has five downloads — more downloads than detail views. That inversion is only possible because the two counters measure different clients: views are page loads by something that renders the site, downloads are file fetches by anything with the URL.
An agent that fetches a skill's zip directly never "views" the page. A crawler that loads the HTML but downloads nothing contributes a view and no download. The two numbers describe overlapping populations without being a funnel.
What the correlation means
At r=0.46, skills that get seen tend to get fetched. That is unsurprising and nearly uninformative — it is the same statement as "exposure works."
The interesting question is what the residuals look like: which skills get many views and few downloads, and vice versa. We do not have a clean enough sample to answer that yet, because both counters are dominated by automated traffic and small numbers.
Why we publish these at all
Because the alternative is worse. A directory that shows only a score and hides the traffic is asking you to trust a number with no context. Showing both invites the reader to see that the score and the traffic are different things — which is the honest lesson.
The same reasoning applies to our own analysis. When we write that score correlates with downloads at 0.47, we should show the medians next to it so the reader can see how thin the underlying data is. A median of two views is not a market.
What would make this analysis better
Three things, none of which we have today:
- Separate human from automated fetches for downloads, the way we now do for page traffic. Without it, every download statistic is a mixed signal.
- Track activation, not just acquisition — how often an installed skill actually fires.
- Longer history. This directory is months old. Most of the curves in it are still forming.
Until then, treat both counters as exposure proxies with wide error bars, and read the score separately.