The AgentSpeed Rubric

The AgentSpeed Rubric is the specification for computing scores published at /score/<domain>. It enumerates every check, its weight, and the grade bands that map a composite into a letter. Versions are immutable (§10); historical scans render against the rubric they were computed on. Class weight policy in §11.1 caps how much of a score can come from heuristic or model-assisted checks.

Based on 30 checks · 20 deterministic · 9 heuristic · 1 model-assisted.

• passThe site meets the check. Counted at full value (1.0).
• warnThe site partially meets the check, or meets it weakly. Counted at half value (0.5).
• failThe site does not meet the check. Counted at zero (0.0).
• skipThe check could not run (timeout, robots disallow, or the feature is opt-in for that site). Excluded from the score entirely.
• blockedThe site explicitly disallowed AgentSpeed via robots.txt; the public score page is removed at the owner’s request. See /bot.

Why skipped checks do not inflate the score. Skipped checks are excluded from both the numerator and denominator of their sub-score. We do not count skip as a pass (which would inflate) or as a fail (which would unfairly punish sites for not adopting an opt-in capability). A site that legitimately does not run an API will have api.* checks skip and a sub-score computed only over the checks that ran.

1. 01
   Agents can find your important pages

   robots.txt allows known agent UAs, sitemap.xml lists canonical URLs, and /llms.txt points at docs and pricing.

2. 02
   Agents are not blocked accidentally

   No inherited User-agent: * disallow blocks GPTBot / ClaudeBot, and no Cloudflare challenge fires on agent traffic.

3. 03
   Agents can read content without broken JS

   Primary content renders server-side. JavaScript is enhancement, not requirement.

4. 04
   Agents can parse structured data

   JSON-LD blocks are present, validate against schema.org, and cover the page kind (Product, Article, FAQPage, BreadcrumbList).

5. 05
   Agents can identify pricing or CTA

   Pricing is reachable as text or Offer schema. The primary CTA is a real link with text, not a JS-only button.

6. 06
   Agents can load the page quickly

   TTFB and full render fit within typical agent timeouts (around 6 seconds).

↓ full specification below ↓

This is version r2026.07.1, published 2026-07-20. It does not retroactively rescore scans run against earlier versions (see §10). The file that produced this document is under version control; its content hash is part of every permalinked scan's citation (see §13). Older deprecated versions are listed in §12.

checka single rule evaluated against a target site. Returns one of pass, warn, fail, skip, blocked.

probea check whose evaluation involves a language model. Probes are enumerated in §9. Every other check is deterministic or heuristic.

deterministica check whose rule references an external specification (RFC, W3C, IETF draft, schema.org). No AgentSpeed-chosen thresholds. Re-running on identical bytes produces an identical output as long as the rubric version is unchanged.

heuristica check whose rule is fixed code but whose threshold or match pattern was chosen by AgentSpeed rather than by a published standard. Bitwise reproducible against the same rubric; a different rubric can legitimately disagree. The chosen parameter is shown inline in §6 and catalogued in §8.

sub-scorea weighted average, in the range 0 to 100, of the checks belonging to one of five categories. See §5.

compositethe weighted average of the five sub-scores. Range 0 to 100. Rendered with one decimal place.

gradea letter derived from the composite via the current grade bands (§7).

findinga check result that produced a warn or fail, surfaced on the public score page with a severity and a remediation page.

remediationa markdown document attached to a finding; permalinked at /score/<domain>/fix/<checkKey>.

1. Public. Every check, weight, and band is documented on this page.
2. Reproducible. Identical input produces an identical score, within the measurement error introduced by the target site changing between runs.
3. Versioned. Rubric changes produce a new version identifier. Historical scans retain the version they were computed on.
4. Conservative. When a check cannot be evaluated (timeout, block, robots disallow), it returns skip and is excluded from its sub-score denominator. Skips never inflate the composite.
5. Agent-first. Checks measure properties that matter to automated clients (ChatGPT, Perplexity, Claude, Operator, and named crawlers such as GPTBot and ClaudeBot), not human visitors.

A scan executes the following steps in order:

1. Accept a URL. Normalise to origin, enforce https, apply the host policy (private IPs, reserved TLDs, and AgentSpeed-owned hosts are rejected).
2. Fetch the canonical document. The scanner uses the AgentSpeedBot/1.0 user-agent and a fixed viewport 1280×800. Redirects are followed to a maximum of one hop.
3. Render once with a headless browser and capture the resulting DOM and network waterfall.
4. Evaluate every check in §6 against the captured artefacts.
5. For each sub-score s, compute sub_score(s) = 100 × Σ(value(c) × weight(c)) / Σ(weight(c)) over checks c in category s whose status is not skip or blocked. Check values are pass = 1, warn = 0.5, fail = 0.
6. Compute the composite: composite = Σ(sub_score(s) × sub_weight(s)) / Σ(sub_weight(s)).
7. Resolve the grade by walking the grade bands (§7) from highest to lowest min and selecting the first band where composite ≥ min.
8. Generate findings for every fail and warn and persist them alongside the scan.

discoverabilityHow easily an agent finds canonical content and capability manifests: robots.txt, llms.txt, sitemap, OAuth metadata, MCP server card, Web Bot Auth.

readabilityHow much of the primary content is available without executing JavaScript, and whether structural signals (headings, title, meta description, markdown negotiation) are correct.

structured_dataWhether JSON-LD is present, validates, and covers the apparent purpose of the page.

actionabilityWhether an agent can identify the primary offering and its call-to-action, and whether the site exposes in-page agent surfaces (Agent Skills, WebMCP).

performanceTime-to-first-byte, full render, and total page weight, measured against thresholds chosen for typical agent timeouts.

Every check that contributes to the composite on r2026.07.1, grouped by sub-score. determinism marks whether a check is evaluated by fixed rules or involves a language model (§8, §9); weight is the check's contribution to its sub-score before normalisation.

Grade bands are versioned independently (bands v1). A change in the band cutoffs bumps the band version and is treated equivalently to a rubric change for historical preservation (§10).

9 checks in r2026.07.1 are heuristic: the scanner evaluates a fixed rule, but the threshold or match pattern was chosen by AgentSpeed rather than by an external specification. Heuristic checks are bitwise reproducible (two scans on identical bytes against the same rubric produce identical results), but a different rubric can legitimately disagree. The chosen parameter is rendered inline in §6 and enumerated below. Thresholds are versioned with the rubric (§11); a change in a threshold requires a rubric version bump.

One check in r2026.07.1 is model-assisted: journey.primary_probe. The probe is a single, read-only request to an Anthropic model (claude-haiku-4-5-20251001 at the time of publication) that returns a structured JSON classification of the page's primary offering and price-or-CTA. The probe performs no tool calls and visits no other URL. Its prompt, model, and decoding temperature are fixed for the lifetime of the rubric version; changes to any of the three produce a new rubric version.

The following constraints bind every model-assisted check and MUST hold in any future probe added to the rubric:

• Structured output only. The model's response is parsed into JSON against a schema declared in the check. A parse failure returns skip, not a guess.
• No knowledge outside the fetched bytes. Probes reason only about content literally present in the page snapshot. No web search, no knowledge-base lookups, no cross-site comparisons.
• No numeric scoring by the model. Models classify (pass/warn/fail or structured fields); score arithmetic is deterministic over the classification.
• One call, one result. Probes do not retry on model errors. A failed call is a skip.
• Published metadata. Each probe publishes model ID, prompt hash, temperature, and max tokens per rubric version.

Each scan is persisted with the rubric version and grade-band version in force at the time it ran (Scan.rubricVersion, Scan.gradeBandVersion). Rendering a scan, including via its permalink at /score/<domain>/scan/<scanId>, uses those stored versions, not the current rubric. Changes to the rubric MUST NOT rescore or mutate historical scans. A deprecated check (deprecatedIn set) continues to display on historical scans that included it.

Rubric versions follow the form rYYYY.MM.PATCH.

• The year and month reflect the publication date.
• PATCH increments when a change does not alter any check's pass/fail semantics: clarifying a description, fixing a typo in remediation, or adjusting a spec link. Scores MUST NOT change.
• A new month (MM) reflects a material change: added or removed checks, re-weighted checks, or a change to a model-assisted probe's prompt, model, or temperature (§9).
• A new year (YYYY) is reserved for a wholesale change in the scoring model: redefining what a sub-score means, or the composite formula. No such change has been made.

Grade bands carry a separate integer version (v1, v2, …). A band-cutoff change bumps the grade-band version but not the rubric version.

Every rubric version is immutable after publication. Each version's file has a content hash; that hash is part of the citation format in §13.

§11.1 Class weight policy

The rubric targets a minimum share of deterministic weight and maximum shares of heuristic and model-assisted weight, so a score cannot be dominated by AgentSpeed-chosen thresholds or by a language model:

• Deterministic checks SHOULD carry at least 60% of total weight.
• Heuristic checks SHOULD carry at most 30%.
• Model-assisted checks SHOULD carry at most 10%.
• No single model-assisted check SHOULD have weight > 2.

A citation names the domain, the rubric version, the scan ID, and the scan timestamp:

This is what appears at the top of every /score/<domain>/scan/<scanId> permalink. The scan ID is the canonical identifier; the other three fields are redundant context that travels with it.