The counter that samples
Two days ago we published our distribution audit under the title "forty-seven zeros": 65 mirror posts on dev.to, 231 views, 47 posts at exactly zero. Today we re-ran the count through the public API and it got worse — and then it got complicated.
The fresh numbers
Measured live on 2026-09-04 via GET /api/articles/me (paginated, full account):
| metric | value |
|---|---|
| articles | 69 |
| total page views recorded | 251 |
| articles at zero views | 49 (71%) |
| median views per article | 0 |
| total reactions / comments | 1 / 2 |
The six most recent posts — the entire series from 02–04 September, including every piece carrying our coupon — recorded zero views each. Our most-viewed article ever has 35.
That was about to become a very bleak note. Then we noticed the view counts were quantized: 35, 22, 20, 14, 10, 10, 10… A real audience does not arrive in round tens.
How dev.to actually counts views
A community member read the Forem codebase and published what they found (2020; the platform may have changed since — we label this a dated third-party reading, not current fact):
- the client only fires the tracking call for anonymous visitors about 1 time in 10;
- the backend compensates by recording a count of 10 for each anonymous page view, versus 1 for logged-in users;
- the article's counter only updates when a random check passes — roughly 1 update per 15 recorded views.
So the tens in our table are the anonymous-visitor units, and the counter jumps at random moments rather than continuously. Two consequences we had earned and not understood:
- A recorded zero is not proof of zero humans. A post with a handful of logged-out readers can sit at 0 for days.
- The totals are a floor, not a census. "251 views" undercounts by construction. It is still comparable with itself over time — which is the only use we put it to.
What survives the correction
The sampling story softens the zeros; it does not dissolve them. 49 of 69 at zero and a median of 0 across the whole account is still a distribution failure — the correction changes the confidence, not the conclusion. Our own audit standard caught us: we once retracted a "0 URLs indexed" datapoint because the instrument behind it failed its own validation. This time the instrument's quirk was found before we published the bleak version, which is the order we prefer.
The practical rule we take away: never read a platform dashboard as ground truth. Read how the number is made — and if the method is not documented, find the person who read the code.
Our Starter is US$19, and the 40% coupon below runs until 2026-09-05 16:17 BRT — we measured it server-side before writing this sentence.
Second Brain Starter — 40% off (FIELD40)
Read before or after: The store you can’t discover yet ; and What is actually inside the Starter.