The load-bearing detail is the denominator. The ratio the author broke out was pageviews per visitor, not per session, and one country row came back at 344 visitors against exactly 344 pageviews [4]. An average of 1.00 in a population where every visitor contributes at least one pageview means every visitor contributed exactly one [24]. Over 90 days, all 344 left from the page they landed on. Real traffic has a tail of people who click a second link, which is the author's stated reason for calling it a script even though he says he cannot prove it [5].
Whatever fetched those pages executed JavaScript. The pageview fires from inside the document, so a client that never runs the bundle never lands in that count at all; the author's own reading is that the traffic ran enough JavaScript to fire a pageview, which is why a server-side user-agent block never caught it [6]. The filter and the fingerprint sit on different layers. One reads a request header. The other is only visible in what the page does after it loads.
The scale of it: 344 plus the 222 US visitors at 1.01 pages each is 566 of the 735 visitors PostHog recorded for the window, or 77% [20]. Organic search's 24 is 3.3% of that same total [21]. The estimate of fewer than 150 real humans puts the human share under 21% [22]. Spread across 38 tools, 24 organic visitors is under one visitor per tool per quarter [23].
The identity bug is a separate failure with a similar shape. `distinctId || "anonymous"` writes the literal string whenever the caller has no id [9], and distinct_id is the key events are attributed on, so every unidentified server event landed on one profile with tens of thousands of events [10]. Funnel steps join on that key. Any funnel with a client step and a server step therefore converted at zero regardless of what users actually did [11], and the phantom inflated Direct as well [12]. The most engaged user in the project was a default argument. The fix parses the cookie posthog-js already sets, `ph_<project key>_posthog`, and pulls distinct_id from it, returning null when there is no key or no cookie header [13]. Returning null is the design: drop the event or mint a per-request id, rather than merge unrelated people under a shared sentinel [14]. That line had been in production for months [15].
The source does not isolate what the JSON-LD, the llms.txt or the markdown twins contributed, because there is no before-and-after to compare against. What it does show is the ceiling with all of them shipped, on queries the author says belong to jwt.io, jsonformatter.org and iLovePDF, domains carrying fifteen-plus years of backlinks [16]. He also found no mechanism by which one visitor could produce a second: one static share image, no linkable results, and five games that kept score in `useState` and lost it on refresh [18].
The 1.00 test transfers if two things hold in your own data. The metric has to be pageviews per visitor, since per-session averages hide the returning human who would push it above 1. And each segment has to be large enough that exact equality is improbable; at 344 that is a finding, at five visitors it is noise. The grouping that produced it took thirty seconds [25].
