Every scraping product markets a success rate. Almost none publishes its own traffic, because the honest version of that table usually says the visitors were machines.
This is thirty days of this endpoint's payment log, read properly for the first time, including the finding that changed a product decision.
Method
Three admin endpoints, one month:
/admin/payment-log?date=YYYY-MM-DD— every paywall hit and settlement, with asyntheticflag for probe traffic./admin/usage-daily— unique client identities (masked subnets and payer wallets), never raw IPs./admin/free-buckets— the free-tier counters.
1,545 non-synthetic events between 2026-09-26 and 2026-10-25, plus twenty-nine per-day identity rolls. Nothing here is sampled or estimated.
What the log says
1,533 wall events. 1,533 of them asked for example.com. Two more asked for example.org. Three asked for this site. That is the entire set of targets for a month: six distinct domains.
The three busiest identities produced 1,085 of those events, and their inter-request gaps give them away:
| Client | Events | Most common gap |
|---|---|---|
2607:fb91:51… | 672 | 1,800 s (389 times) |
2001:1708:26… | 247 | irregular, 2,530–4,673 s |
74.220.49.20… | 166 | ~5,410 s |
Nobody types a URL into a terminal every thirty minutes for twenty-eight days. These are uptime monitors — a directory's or the operator's own — polling the endpoint to see whether it is alive. Hour-of-day is flat: between 76 and 123 events in every one of the twenty-four hours, which is automation, not business hours.
Of 98 distinct client identities, 74 appeared exactly once.
And of the twelve settlements the log recorded, all twelve were example.com, example.org or this site — the paid canary and manual tests. Customer revenue in the window: zero, and the run that "found" money had been counting the canary.
What the numbers could not say, and why that was a defect
The log recorded no user agent. A monitor, an agent library and a person were indistinguishable in the data, so a health check read as demand — including inside this company, where it briefly justified building a card checkout for a stream of example.com requests.
That is now fixed at the source: every 402 records a coarse ua_class — monitor, browser, agent_lib, unknown, or internal for the operator's own tooling — alongside the first 120 characters of the User-Agent. Monitors are tested first, because an uptime service is perfectly happy to send a Mozilla prefix. The next read of this table can say who, instead of inferring it.
The second fix is arithmetic: settlements whose target is example.com, example.org or this site are counted as internal, not as revenue. A health check working is not a sale.
What the finding changed
A card checkout was scoped and then put on hold. Not because card processors refuse the category — read their published lists and the ban is narrower than the folklore — but because there is no human at the wall to sell to. Building checkout for a thousand example.com probes a month would have produced a beautifully engineered funnel converting nobody.
The gate that would unpark it is measured, not argued: ten or more distinct real target domains from five or more non-monitor clients in fourteen days, or three humans asking how to pay by card. Today's reading is six domains in thirty days, so the plan waits in a file instead of a sprint.
The three numbers worth tracking instead of hits
- Distinct real target domains per week. Six in a month is the demand number, and it is the only one that has ever been near zero while every other counter looked healthy.
- Distinct non-monitor clients. Because two monitors firing on a schedule can out-count forty humans who each try once.
- Settlements that were not the canary. The only figure that is money.
The uncomfortable part, stated plainly
This is a post about a month in which the product's traffic was almost entirely machines checking whether the product was running, published on the product's own blog. It is being published because the alternative — quoting a success rate with the failures stripped out, and calling a monitor's poll a customer — is what every benchmark in this category already does, and the point of this endpoint is that it tells you what actually happened.
The next one of these will have thirty days of ua_class in it. If the numbers are still six domains and a canary, it will say that too.