Why four scores instead of one
Every agent task decomposes into the same sequence: find the site, interpret what it offers, invoke an applicable operation, and decide whether the result can be relied on. A site can excel at discovery and still be unusable because nothing is callable. One blended number cannot express that difference, so Santos scores each layer separately from 0 to 100 and reports the composite alongside the parts.
Discoverable: can an agent find you
This dimension measures machine-facing entry points: whether llms.txt exists and orients usefully, whether robots.txt is deliberate rather than accidental, whether sitemaps are current, whether canonical URLs resolve consistently, and whether documentation and interface descriptions are linked from public pages and headers. A common failure is documentation that exists but is unreachable by machine: no sitemap reference, no llms.txt, and a robots rule that quietly blocks it.
Understandable: can an agent interpret you
This dimension measures meaning: semantic HTML, JSON-LD identity for the organization, website, and services, consistent product naming, visible pricing, and descriptions that agree across pages. An agent that cannot tell the provider from the product will misquote both, and a pricing page that contradicts the structured data produces exactly the kind of inconsistency an autonomous system should refuse to act on.
Callable: can an agent act
This dimension measures invocable surfaces: OpenAPI documents, MCP servers, capability manifests, typed input and output schemas, stable error models, and x402 payment terms with explicit prices. It applies only where the business genuinely offers operations an agent should invoke; a documentation site with nothing to call is not deficient for lacking an API.
Trustworthy: can an agent rely on you
This dimension measures operational evidence: HTTPS configuration, security headers, accessibility, performance, browser errors, reachable support channels, and whether published claims match observed behavior. Trust is what makes the other three dimensions safe to use unattended, because a callable interface on an unverifiable site is a liability rather than a capability.
Applicability-aware scoring
Santos classifies the target before scoring it. Categories a site does not need are marked not applicable and removed from the denominator instead of being scored as failures, so an informational publisher is not penalized for lacking MCP or payment surfaces while an API provider is held to the full bar. All numbers come from deterministic published rules; AI-generated summaries, when requested, annotate the report but never change a score.
The highest-impact fixes per dimension
Discoverable: publish a concise llms.txt, align robots.txt with intent, and ship a current sitemap. Understandable: add JSON-LD Organization and WebSite markup and make pricing visible and consistent. Callable: publish a valid OpenAPI document or a capability manifest with typed schemas, stable errors, and explicit prices. Trustworthy: fix missing security headers, remove console errors, and expose a reachable support path. In audit after audit these basics move the score more than any exotic addition.