What actually reads this site

The split, with and without one bad hour

11 days of detailed data · to 23 Aug 2026 · sampled, estimates not talliesShare of page views. A page view is a page-shaped path answered 2xx or 404, eyeball traffic only, this project's own tooling excluded.

The five rows split every page view by what the client claimed to be. Each row shows the share without 2026-08-22 first, because that is the primary window, and the share with it underneath, dashed, as a warning rather than a finding. The AI row is the live demonstration, because one bad hour moves it from 3.5%to 16.4%. The last row is everything my classification list does not recognise, and I never present it as people.

The same figures are published as JSON and CSV, so you can check any number on this page against them.

The hour that poisoned the chart

11 days of detailed data · to 23 Aug 2026 · sampled, estimates not talliesShare of page views. A page view is a page-shaped path answered 2xx or 404, eyeball traffic only, this project's own tooling excluded.

On 2026-08-22, 557 requests arrived claiming 7 crawler names that belong to 5 operators, being OpenAI, Perplexity, Google, Anthropic, Amazon. 539 of them landed in a single hour, every request came from one country, and 95.0% were answered 404, meaning they asked for pages this site has never served.

The obvious question is who sent them. On the evidence I keep, spoofing is the near-certain reading and cannot be proven, because proving it needs the IP addresses behind those requests, and I refuse to hold them. No operator named here is accused of anything. The names were the costume, and an operator whose name was worn is the party imitated, not the party acting.

That single hour multiplied the apparent AI share 4.7 times. Any chart that trusts user-agent strings can be moved the same way, by one actor, in one hour.

What the unrecognised traffic claims to be

11 days of detailed data · to 23 Aug 2026 · sampled, estimates not talliesClaimed Chrome major versions among unrecognised page views. Each bar's denominator is the Chrome-claiming slice, named beside the chart.

Unrecognised means my classification list matched no known client, and that is all it means. It is not evidence of a person, which is why the largest slice of this site's traffic is labelled "Unrecognised clients, not verifiable as people" and never anything warmer. This section is what that label links to.

84.3% of unrecognised page views claim to be Chrome. Of those Chrome claims, 43.9% name major versions released before 2024. The two figures have different denominators, so they stay in prose and never share an axis. Chrome 120 went stable on 5 December 2023 and the next major not until 23 January 2024, so a claim of major 120 or older is a claim of a pre-2024 browser.

A client claiming Chrome 78, released 22 October 2019, appeared on 11 of 11 observed days.

Release dates come from the Chromium Dash release schedule, the schedule service the Chromium project itself runs, and each date in the table links to the exact record. The dashed bar is a warning, because that spike is largely two burst days, not a finding about ordinary traffic.

Most page-shaped requests ask for pages this site never served

11 days of detailed data · to 23 Aug 2026 · sampled, estimates not talliesShare of page-shaped requests answered 404, three windows, each named beside its figure.

A 404 on this site returns a real HTML page, so my page-view rule counts a page-shaped request answered 404 as a page view. The rule was set by checking it against Cloudflare's own page-view total for the same days, because counting successes only would miss most of what actually asks for pages here. Most of the 404 numerator comes from unrecognised clients, so this is a claim about request behaviour, never about who is behind it.

Both bars are honest findings with stated windows, so neither is dashed. 2026-08-23 carried a second, smaller burst of requests for pages that never existed, which is why the table also names the share without both days.

Who actually reads the markdown mirror

10 days, without 2026-08-22 · to 23 Aug 2026 · sampled, estimates not talliesShare of markdown mirror fetches by classified client bucket. My own fetches are excluded.

Every page here has a markdown twin at the same address ending in .md, published for machines. Three limits sit on this chart, in the open. The AI slice rests on a small number of fetches and I hold it at first-signals strength. A client that never discovered the mirror could not have read it, so shares describe observed readers, not appetite. And my own fetches are excluded before anything is counted.

ClaudeBot is the wrinkle worth naming. In this window the markdown mirror made up 45.7% of its own content fetches, so nearly half of what it took from this site was the mirror rather than the HTML.

I publish a markdown mirror of this page too, at /visibility.md, so a later window can include the readers of the page you are reading now.

Robots.txt etiquette, per crawler

10 days, without 2026-08-22 · before the 2026-08-24 policy changeRobots.txt fetches per unit of content taken, both denominators shown for every crawler.

Both columns divide the same numerator, a crawler's robots.txt fetches in this window. The first divides by the HTML pages that crawler took. The second divides by its content fetches with the markdown mirror counted as content, which is the stricter reading. Showing only the larger ratio would be a cherry-pick, so every row carries both. This is a table rather than a chart on purpose, because an observed zero deserves words, never a zero-length bar.

Every name here is self-identified, and etiquette observed is not obedience proven. On 2026-08-24 I changed the site's robots policy, so this table is the before-baseline, and a later refresh adds the after columns beside these.

What Google shows, and what it withholds

21 settled days · to 18 Aug 2026 · Google Search ConsoleImpressions whose query Google named, over all impressions, settled days only.

Across 21 settled days, Google named the search query behind 2 of 245 impressions, which is 0.8%. The rest are queries Google declined to name. An impression is one appearance of this site in someone's results, and a day only counts here once Search Console has stopped revising its figures, which is what I call settled.

The withholding is a known mechanism rather than something odd about this site. Ahrefs measured it at 46.77% across the sites it studied, so a site this small reads as the far end of the same mechanism, where almost every query disappears. That external figure stays in prose and never enters a chart or table beside this site's own. The interesting number is not the share but how it moves, so the table below gains a row at each refresh and only ever holds windows that already exist.

First signals, and what I am holding back

First signals are observations too few to claim a pattern, published with their size stated so nobody mistakes them for one. At launch, exactly one figure on this page is held at that strength, the AI slice of the markdown mirror's readers above, and nothing else here is built on it.

The findings below are withheld on purpose, which is itself a statement about method rather than a promise of coming content. Each row names the condition its figure waits on, and a row leaves this list only through a review round.

• llms.txt before and afterThree to four weeks of after-data; the before-baseline and go-live date are already recorded.
• Crawl versus credit, Google editionAll joint Search Console days settled.
• AI training / search / user-fetch splitAI page views in the hundreds; one crawler session currently moves the split by 24 points.