// geo-standard@2026.05

Methodology

How we compute our agent-readability score, what “pass” means for each signal, how often we refresh the dataset, and where the measurement stops being precise — laid out so anyone can reproduce it.

The five signals

Each domain is scored on five signals, weighted to a 100-point total:

  • Semantic HTML — 25 pts. Landmark elements, single H1, alt text coverage.
  • JSON-LD — 20 pts. Organization at root; type-specific schema per leaf (Article, Product, FAQPage, BreadcrumbList).
  • llms.txt — 15 pts. Presence, validity, and content coverage at /llms.txt.
  • Citability — 20 pts. Front-loaded answer, named entities, numbers and dates in the first 50–70 words.
  • Speed — 20 pts. TTFB under 800ms (full 100 below 300ms); SSR HTML below 1MB.

Pass thresholds

A site "passes" a signal when it clears roughly 75% of that signal's max — the empirical line above which AI engines reliably extract a clean answer:

  • Semantic ≥ 19 / 25
  • JSON-LD ≥ 15 / 20
  • llms.txt ≥ 11 / 15
  • Citability ≥ 15 / 20
  • Speed ≥ 15 / 20

Total-score milestones: ≥ 85/100 "agent-native" (AI engines cite by name); < 55/100 "effectively opaque" (engines time out or skip).

Dataset construction

The Agent Readability Leaderboard tracks AI-industry companies across four canonical categories — Infra, Models, Agents, Dev Tools. Flagship rows (top ~30) are hand-scored; long-tail rows are deterministic estimates seeded from domain characteristics and re-scored weekly by an automated scan against the same five-signal scorer.

Any row can be re-scored live by anyone at /check?u=<domain>. Discrepancies between the leaderboard score and a live re-scan are expected when a site ships changes between scheduled scans.

Scan cadence

Weekly re-scan of the full dataset via /api/public/hooks/rescan-leaderboard. Quarterly report aggregates the most recent four weekly snapshots. Monthly data drops surface single-stat findings between quarterly cuts.

Limitations

  • Marketing pages only. Docs sites and product UIs are out of scope — their access patterns and citation needs differ.
  • Public surface only. Auth-gated content cannot be scored; sites that hide everything behind login will under-score regardless of internal quality.
  • Snapshot in time. Scores reflect the most recent scan; sites that ship daily will drift between scans.
  • Five signals, not all signals. Backlinks, brand authority, and Information Gain affect citation rates but are out of scope for this measurement.

License & attribution

Report text, headline statistics, and the underlying dataset are licensed CC BY 4.0. Attribution: "grow.contact, State of the Agent-Readable Web (CC BY 4.0)" with a link back to the report URL.

Dataset endpoint: /api/public/leaderboard.json

Changelog

  • geo-standard@2026.05 — Current. Adds Citability weighting refinement (front-loaded answer detection).
  • geo-standard@2026.02 — Added speed threshold step function below 300ms TTFB.
  • geo-standard@2025.11 — Initial five-signal scorer released.