crail
ARTICLE

Introducing Crail's agent-readiness score

A 0-100 score, published with its full methodology, measuring how usable a piece of software actually is for an AI agent to discover, evaluate, and buy.

Crail Editorial · 2026-07-10productmethodology

Star ratings measure whether people like a product. They don’t measure whether an AI agent — increasingly the thing doing first-pass vendor research on a buyer’s behalf — can actually reach, understand, and transact with it. That’s a different question, and as of this launch, Crail publishes an answer to it on every vendor page: the agent-readiness score.

What it measures

The score is a simple average of six sub-scores, each 0-100:

  • Discoverability — does the vendor publish an llms.txt, does robots.txt allow known AI crawlers, is there schema.org markup on their site?
  • Machine-readable content — are docs clean markdown/text, or a JS-only single-page app that a scraper sees as a blank shell?
  • Programmatic access — can a developer self-serve an API key and sandbox without a manual approval step?
  • MCP/agent-native support — does the vendor publish and maintain a real MCP server (verified by us actually connecting to it, not just a claim)?
  • Self-serve transactability — can you buy the product without a human sales call?
  • Transparency — is pricing fully public, is there a public status page and security docs without an NDA?

What it’s not

It is not a quality score. A vendor can build an excellent product and score low here simply because their pricing is quote-only and their docs sit behind a JS wall — and a mediocre product can score well by being well-instrumented for agents. We publish both agent-readiness and pricing/feature data side by side specifically so neither gets mistaken for the other.

It’s also not for sale. Vendor subscriptions on Crail (Crail Verified/Reach) pay for data completeness, freshness, and demand-gen — never for a higher score. The scoring code path and the billing code path don’t share a database field, on purpose.

Methodology version

This is crail-ar-v0.1. When the methodology changes, the version number changes with it — old scores stay attributed to the version they were actually computed under, rather than silently drifting. Full breakdown at /methodology.