Retrieval Benchmark

How Wheel of Heaven measures AI information penetration — the quarterly benchmark of whether the corpus appears in AI answers' retrieval and citation chains.

≈ 3 min read 459 words Updated 2026-08-23

On this page

Wheel of Heaven’s distribution goal is not pageviews but pathways: the number of information routes that eventually traverse the corpus — including answers a reader never sees the source of, because an AI assistant retrieved and synthesized it. The retrieval benchmark measures progress toward that goal.

What it measures#

Information penetration — the share of relevant informational questions for which Wheel of Heaven appears anywhere in an AI answer’s retrieval/citation chain. This is the project’s north-star KPI: pathways traversed, not pages viewed.

Method#

  • Cadence: quarterly (Feb / May / Aug / Nov). Each run is a dated file in the project’s plans/benchmark-runs/.
  • Surfaces: Google web search (an automatable proxy for RAG retrieval), plus ChatGPT, Perplexity, Gemini, and Claude with web search. Manual surfaces get one pass per run; each question is asked verbatim, no follow-ups.
  • Recorded per (surface × question): whether the corpus was cited, which page, position among citations, competing sources, and whether the corpus’s position was represented accurately.
  • Scores per run: penetration % per surface and overall; accuracy % among citations; a competitor-frequency table (who wins where the corpus doesn’t).

The question set is fixed and versioned. Questions are asked as written — no “wheelofheaven” hints except in the project-direct category, which measures brand retrieval rather than topical retrieval — and are never tuned toward known strengths between runs. Additions get a version stamp; questions are never silently removed.

Question set (v1 — 50 questions, 7 categories)#

CategoryCountMeasures
Canon-internal8Raëlian doctrine, the Elohim, the embassy
Comparative mythology12Flood myths, divine councils, Watchers/Nephilim
Precession / world ages6Ages, the Great Year, Hamlet’s Mill
Primary texts8Gilgamesh, Enûma Eliš, Enoch, Theogony
Authors / figures6Sendy, Sitchin, von Däniken, the plural Elohim
Structured data5Machine-readable / CC0 comparative-mythology data
Project-direct5Brand retrieval for “Wheel of Heaven”
CategoryPresentPenetration
Project-direct4 / 580%
All others0 / 450%
Overall4 / 508%

The baseline’s findings shaped the subsequent work:

  • Brand retrieval resolves to the GitHub org and the legacy .io domain, not .world — only one .world URL surfaced in the entire run. The .io GitHub Pages source was subsequently neutralized to a redirect.
  • Dataset queries return nothing that serves the intent — from anyone. The comparative-mythology datasets and their landing pages were aligned with the query language the benchmark used.
  • Wikipedia is weaker competition than expected; the recurring competitors are the Christian-apologetics cluster, the Heiser divine- council ecosystem, and fringe/new-age precession sites.
  • Several open lanes with no credible competition (Council of Eternals, Watchers-vs-Nephilim, the twelve world ages) are candidate content targets.

Movement in the comparative and primary-text categories over successive quarters is the signal that the knowledge-system strategy is working.

Edit this page on GitHub