# How the AgentCapable scan measures

Two questions decide everything: can agents get in, and can they read you
once they are in. Machine-readable check catalog: https://agentcapable.dev/api/checks

## The instrument (roster roster-v1.0)

Your site is requested with the exact User-Agent strings real AI agents send,
from two independent datacenter networks, beside three controls (a normal
browser, a nonsense UA, Googlebot).

- Training crawlers: GPTBot, ClaudeBot, CCBot, Meta-ExternalAgent, Bytespider
- AI search indexers: OAI-SearchBot, PerplexityBot, Claude-SearchBot, Applebot
- Live assistant fetchers: ChatGPT-User, Perplexity-User, Claude-User
- Policy tokens (read in robots.txt, never requested): Google-Extended, Applebot-Extended
- Controls: Chrome-Control, Nonsense-Control, Googlebot-Control

## The verdict

Five headlines: Open, Partially blocked, Agent-hostile, Contradictory,
Unknown. Rules: (1) a finding counts only when BOTH networks saw the same
thing; (2) a declared robots.txt block that the edge enforces exactly gets
full credit (we measure policy-enforcement agreement, not openness); (3)
what we cannot measure is excluded and labeled Unknown, never guessed.

## The score (rubric rubric-v1.1)

score = 100 x passed-weight / decided-weight over the 23 checks below.
Unmeasurable or not-applicable checks drop out of the denominator. Fewer
than 5 decided checks renders "?". No LLM touches the score. Weights are
stated, versioned expert judgment; recorded runs from the paid Live Agent
Test are archived and aggregate evidence will inform future rubric versions,
always as a new version number, never a silent change.

### Access

- **robots.txt is readable and sane** (AC-ROB-02, weight 12): robots.txt is the first file every agent reads; if it is broken or self-defeating, everything downstream suffers.
- **Agents can get in** (AC-UAD-02, weight 12): This is the bottom line: when an AI agent knocks, does it get the page?
- **Agents get the same content as a browser** (AC-UAD-03, weight 6): Some sites let agents in but serve them a stripped-down page; the agent then answers from thin content.
- **No page-level AI bans (noindex / noai)** (AC-ROB-04, weight 5): A single meta tag can quietly remove your pages from AI answers.
- **Live assistants are not silently denied** (AC-ROB-03, weight 3): When a person asks their AI about your site mid-conversation, this is the request that fails.
- **No CAPTCHA on the front door** (AC-ROB-06, weight 3): Agents cannot solve CAPTCHAs; one on the front door turns every agent away.

### Discovery

- **Sitemap exists and is valid XML** (AC-DIS-01, weight 5): The sitemap is how agents and crawlers enumerate what you publish.
- **Agent files served with correct content types** (AC-DIS-04, weight 2): A file served with the wrong type gets misread or discarded by strict clients.

### llms.txt

- **llms.txt exists** (AC-LLM-01, weight 8): llms.txt is the welcome note AI systems look for first; sites that have one get summarized on their own terms.
- **llms.txt starts with a title** (AC-LLM-02, weight 5): Without the title line, many parsers treat the file as malformed.
- **llms.txt fits an AI's reading budget** (AC-LLM-03, weight 3): Agents read with a fixed budget; an oversized file gets truncated mid-sentence.
- **llms.txt links actually resolve** (AC-LLM-04, weight 3): Agents follow these links literally; each dead one is a dead end in a conversation.

### Readability

- **Content is readable without JavaScript** (AC-REN-01, weight 10): Most agents read raw HTML and never run your JavaScript; an empty shell means you are invisible.
- **Markdown when asked (content negotiation)** (AC-REN-02, weight 8): Markdown is roughly 10x cheaper for an AI to read than rendered HTML.
- **Missing pages return an honest 404** (AC-REN-09, weight 5): A page that returns 200 OK while actually showing a not-found message sends agents in circles.
- **No login, consent wall, or paywall in front of content** (AC-REN-10, weight 5): Whatever sits in front of your content is all an agent will ever see of it.
- **Clean status and redirects** (AC-REN-11, weight 5): Redirect chains and cross-host hops make agents abort the fetch.
- **A markdown mirror is fetchable** (AC-REN-03, weight 3, dev-platform/docs only): A markdown twin gives agents a clean, cheap version of each page.
- **Page weight fits an AI's budget** (AC-REN-07, weight 3): Oversized pages get cut off mid-read; whatever falls past the cut is never seen.
- **Markdown alternate is declared** (AC-REN-04, weight 2, dev-platform/docs only): The declaration is how an agent discovers your markdown version exists.
- **Forms work without JavaScript** (AC-REN-12, weight 2): A form without action and method attributes only works with JavaScript, which most agents never run.

### Structured data

- **Structured data (JSON-LD) present** (AC-SD-01, weight 4): Structured data lets agents quote your facts precisely instead of guessing from prose.

### API

- **OpenAPI spec published (developer platforms)** (AC-API-01, weight 5, dev-platform/docs only): An agent integrates with a developer platform through its API; the published spec is how it learns the endpoints.

## What we never claim

Requests come from datacenter networks, not AI vendors' verified addresses.
The strongest claim a report makes: your edge treats agent requests from
cloud servers as hostile. Never that a specific vendor's bot is blocked.

## Versioning

Check ids are stable. Weight changes arrive only as a new frozen rubric
version (hash-guarded in CI). Gate CI on ids from /api/checks.