Machine visitors versus human visitors on six small sites, measured

Axion Labs · getaxionlabs.com · read 17 September 2026 · next read 14 October 2026

Each of our sites counts its visitors twice. Every request is bucketed by the operator its User-Agent declares and by what it asked for, so a machine reading an answer is separated from one probing for a login form: the path classes are machine, page, crawl, probe and beacon. Human pageviews are counted separately, by a script that only runs in a real browser. The table is both meters over one rolling 30 day window, read 17 September 2026, on the six smallest sites we run.

SiteDeclared agent requestsHuman pageviewsMachine to humanTop agent class
appealsdesk.co.uk4,3201,6952.5 timesOpenAI
denialfacts.com4,5981,8372.5 timesOpenAI
nycticketfacts.com2,8067983.5 timesOpenAI
screeningfacts.com2,5745994.3 timesAnthropic
mtdfacts.co.uk1,4727352.0 timesAnthropic
open.getaxionlabs.com113none recordedno human sideOpenAI
All six15,8835,6642.8 timesOpenAI

Two other metered properties are left out because they serve machines only: our MCP gateway, which took 24,063 declared agent requests in the same window, and a books subdomain that recorded nothing.

Every machine number is a floor

The agent column counts declared User-Agent strings. An agent arriving in an ordinary browser User-Agent is invisible to it, as is anything reading us through somebody else's cache. The human column has the opposite blind spot: a person reading with scripts turned off is never counted. That is why we publish both counts beside the ratio. Read every machine figure here as a floor, never a ceiling, and the ratios as indicative rather than exact.

How to repeat this

Bucket every request your server answers by the operator named in its User-Agent, keeping the path so a machine reading a page is not mixed in with one probing for a login form.

Count human pageviews over the same window with a small script in the page, because a script runs for a browser and not for a crawler.

Divide the first by the second, per site rather than across a whole estate, and publish both counts and the window beside the ratio so a reader sees what each number misses.

The tool, and the door

The reader behind this table is public: github.com/endrezsoltdios-sketch/axion-readers. It needs no analytics vendor and no key. For the raw monthly figures behind this table, or to have us run the same reading against your sites, write to hello@getaxionlabs.com. That door is ours and a person answers it.