Hekhal · measurement · first edition

110 sources answer for the mystical traditions. 96 of them appear once.

When someone asks an assistant what the Zohar is, or what the Watchers did, or what Dionysius meant by darkness, the answer is assembled out of whatever the machine can reach. We measured what it reaches for. The head is four or five reference works. Behind them is a long tail of sites that each appeared exactly once.

The shape of it

Every domain that answered more than one of the 12 questions, by how many it appeared in. Bar length is the count; colour marks what kind of source it is, and each bar carries its number so the colour is never doing the work alone.

• Encyclopedia or reference work · 53 links
• Primary-text library · 43 links
• Journal, press or university · 32 links
• Forum, blog or upload · 15 links
• Other site · 54 links

And the tail, at the same scale

Each mark is one domain that appeared in exactly one answer. This is the same measurement as the bars above, drawn as mass rather than as a list, because the shape is the argument.

96 domains, one appearance each.

One corpus, two engines, opposite answers

The same site, the same subjects, measured on two engines in the same period. Microsoft's figure is from Bing Webmaster Tools, which reports what Copilot cited. The Perplexity figure is this study. The other assistants are not drawn as zero, because we have not measured them.

Why the gap exists, as far as we can tell

The obvious explanation is that the assistant cannot reach the site. That is measurably not the case. Hekhal's robots.txt permits every AI agent by name, and in the seven days to 2026-09-20 Perplexity's crawler fetched this site 246 times.

It reads the site thoroughly and rarely surfaces it. The difference tracks what each engine retrieves from. Copilot leans on Bing, where this property is verified and its sitemap was submitted on 2026-09-01. Perplexity's retrieval tracks ordinary web ranking, and in Google Search Console this site sits at position 60 to 78 on exactly these subjects.

Access was never the constraint. Retrieval rank is. That claim is falsifiable and we intend to test it: if the rank on one subject improves, citation on that subject should follow, and if it does not, this paragraph is wrong and we will say so here.

The distance between how much the answer layer uses a corpus and how much the click economy rewards it is what the sibling network calls the click gap; the term and its rationale are theirs, on the Lattice geo-transparency page.

Every question, and what answered it

The twelve questions, each with the number of sources displayed, how concentrated the answer was on its single largest source, and whether the primary texts were represented at all.

The most fragmented answer was what is the book of the watchers in 1 enoch, where 15 links came from 15 different domains. The most concentrated was where can i read the zohar in english for free, where sefaria.org supplied 42% of everything shown.

What this does not show

If this is your field

Rerun it, or extend it

Both files are CC-BY 4.0. The protocol names the queries, the capture method, the normalisation rule and the honesty constraints, so the study can be repeated by someone who does not trust us, which is the only kind of repetition worth anything.

• protocol.json · the queries and the method
• perplexity-2026-09-21.json · the raw capture

Corrections and contrary results are welcome at the editor, and are recorded on this page rather than folded in silently. This is the first edition; each rerun is published beside it rather than replacing it, so the series can be read as a series.

By subject