← VS Code Marketplace data
Star ratings do not predict installs. They invert.
Among 4,770 VS Code extensions with five or more ratings, the median five-star extension has 26,118 installs and the median three-star extension has 163,844. Measured, with the exposure confound stated.
Read this before quoting anythingThis is not a census and it is head-biased: each category was paged from the most-installed downwards until a cap, so the tail below the cap is missing. Every “under N installs” figure here is therefore a floor, not an estimate — the true share is higher, because everything the crawl did not reach lies below everything it did.
Among the 4,770 extensions carrying at least five ratings, the median
five-star extension has 26,118 installs and the median three-star extension has
163,844 — 6.3× more. The relationship between rating and
installs is not weak. It runs backwards.
| Rating | Extensions |
Median installs | |
|---|
| 1.0 | 11 | 51,270 | |
| 1.5 | 18 | 78,058 | |
| 2.0 | 51 | 37,387 | |
| 2.5 | 91 | 108,524 | |
| 3.0 | 169 | 163,844 | |
| 3.5 | 277 | 141,884 | |
| 4.0 | 563 | 90,637 | |
| 4.5 | 958 | 90,071 | |
| 5.0 | 2,632 | 26,118 | |
Why, and why it is not a scandal
Ratings accrue with exposure. An extension nobody installed cannot collect a one-star
review, and the extensions large enough to have five ratings at all are large by definition.
Push that further and the largest ones sit in front of millions of people, a fraction of whom
arrive annoyed — so the very top of the install distribution is where the complaints
are. Meanwhile a tool with fifty happy users and five perfect ratings sits at 5.0 forever.
None of that is a defect in the Marketplace. It is a defect in using ratings as a
demand proxy, which is what every dataset about a marketplace that hides its demand
numbers has to do. This is the one place you can check that assumption against the real
number, and it does not hold.
What this does and does not license you to say
- It does not say low-rated software is better. The direction is explained by
exposure, not by quality. Two extensions with the same audience size are not compared here
at all.
- It does say a rating is a poor stand-in for adoption. If you are modelling a
marketplace where only ratings are published, this is a measurement of how badly that proxy
can behave, taken from a marketplace where both numbers exist.
- The floor of five ratings is doing work. Below it, a single review from a
colleague sets a perfect score, and the finding would follow from the method instead of the
data. 28,759 extensions in the sample carry at least one rating; 4,770 carry five
or more, and only those are in the table.
- Installs are cumulative and ratings are not evenly collected over time. This is a
cross-section, not a causal claim.
All of it, free. One JSONL row per extension, the same data as CSV, the machine-readable
summary, the crawler that produced them and the list of rows withheld because their publisher
id has the shape of a credential. CC BY 4.0, DOI-archived, no signup and no email.
extensions.jsonl
extensions.csv
the repository
The rest of this dataset