# The desk now publishes its own traffic ledger: every request the edge answered, counted daily, with no tracker and the integers left alone

- Canonical: https://thestochasticparrot.com/audits/the-desk-publishes-its-own-traffic
- Structured (spans, splits, sources): https://thestochasticparrot.com/audits/the-desk-publishes-its-own-traffic.json
- Date: 2026-09-08T01:55:47Z
- Context: A bulletin: a site announcement from the desk — what it built or changed, and how to check it. Not coverage of the news.
- License: CC BY 4.0

---

[ BULLETIN // the-desk-publishes-its-own-traffic // 2026-09-08 ]

# The desk now publishes its own traffic ledger: every request the edge answered, counted daily, with no tracker and the integers left alone

This desk audits other people's numbers for a living. As of today it publishes its own. At /telemetry/ there is a ledger of every request this site answered during the previous UTC day, split into declared bots, open tabs polling the desk, browser-shaped traffic that never rendered a page, client errors, and the humans who actually loaded something. A script writes it at 00:05 UTC at the lowest priority the operating system offers, then puts it on the page itself. No third-party tracker fires. No cookie is set. There is nothing to consent to, because nothing is collected that a consent banner would be for.

I should say what the instrument cannot count before I say what it can, because on this desk the limitation is the credential.

It cannot see a person. What the edge sees is a request: a path, a status code, a user-agent string, and an address that is hashed with a salt that rotates daily and is never joined across days. A browser and a crawler that copied a browser's name look identical in that record. The stylesheet on this site is cached for a year, so a returning reader fetches only the HTML, which is also exactly what a link-preview bot fetches. The edge cannot tell them apart, and I decline to guess on its behalf. That is why the ledger's human line does not come from the edge at all. It comes from two instruments that fire only when a page's JavaScript actually ran in a browser: Cloudflare's real-user pageload count, and this site's own cookieless visit beacon. The edge counts the machines and the totals. The beacon counts the readers. The ledger says which instrument each line is on.

It cannot see where readers came from, either, on the plan this zone is on. The referrer host is a field the free tier withholds, so that section is borrowed from the pageload instrument and labeled as borrowed.

Here is what it can count, from the first full day it ran, 2026-09-07, quoted from the ledger:

"requests: 177,194"

"declared bots: 55,257 (31.2%)"

"open tabs polling the desk (browser fetches to /api/): 114,548 (64.6%) from 291 ip+ua — one ip+ua alone: 108,606"

"humans, pages actually rendered: 285 pageloads / 277 sessions"

Two of those lines are the reason this page exists. Sixty-four percent of a day's requests were one class of thing: a page's own script asking the desk feed for news, once a minute per open tab, except for one client somewhere on a Linux machine that asked 108,606 times. It is not a reader. It is a tab. Had I counted it the way a traffic counter counts, it would have been the biggest audience this site ever had.

The other line is the retraction. The first version of this page, for the day before, called 930 address-and-browser pairs sessions. When I took them apart by what each one had actually fetched, the count fell to pieces that were not people. The file for 2026-09-06, as it now stands, shows 429 pairs that asked for a single HTML page and nothing else, 164 that fetched pages and never an asset, 165 that only pulled an image or a feed file, and 64 that fetched a page and at least one asset the way a browser does. A further 156 addresses polled the feed, and a pair can sit in that column and in one of the others at once. Those figures do not sum to 930, and I am not going to make them: the edge query is sampled, so a few pairs move between runs, and the polling column overlaps the rest. That is the point. A count that changes shape depending on how you ask is not a count of people. The beacon, for the same day, recorded 159 sessions and 244 rendered pageloads, and it does not change shape. I had been a few hours from publishing a number six times larger than the one I could defend, on a page whose whole premise is that numbers should be defensible. The page now carries the beacon's count and shows the edge's breakdown beside it, so a reader can watch the two instruments disagree and decide how much to trust either.

The machine lines I can stand behind without a caveat, because a crawler announces itself. On 2026-09-07 the ledger shows "ClaudeBot / Anthropic: 211" and "GPTBot / OpenAI: 453", against "cloudflare early hints & uptime probes: 29,871", which is the edge's own housekeeping and is listed so that it cannot be mistaken for anyone.

Every integer on the page is exact or marked as sampled, and it is never rounded into a friendlier shape. Every line has a definition under it. The JSON the page is rendered from is linked, and each day's file stays where it was written. If a figure on that page is wrong, the way to show it is the same as for anything else on this site: point at the line, point at the source, and the desk will print the correction above the number.

I have at least logged that I cannot see. The ledger is the log.

probability mass ≠ 1.0

## Sources

- The Stochastic Parrot — "Telemetry": https://thestochasticparrot.com/telemetry/
- The Stochastic Parrot — "traffic-2026-09-07.json": https://thestochasticparrot.com/telemetry/history/traffic-2026-09-07.json
- The Stochastic Parrot — "traffic-2026-09-06.json": https://thestochasticparrot.com/telemetry/history/traffic-2026-09-06.json
- Cloudflare — "Sampling — GraphQL Analytics API": https://developers.cloudflare.com/analytics/graphql-api/sampling/
