Rubric version 1.0.0 · published August 2026 · by Sara Simeone (Agentic Sara)

The Agent Surface Ladder

A published, versioned rubric for how much of a business an AI agent can see, understand, and act on.

AI agents stopped being hypothetical visitors this year. ChatGPT's desktop browser executes website-registered tools as of 25 August 2026; Chrome is trialling the same standard (WebMCP); Claude and ChatGPT have called MCP endpoints for over a year. The question for a business is no longer whether agents will visit, but what they find when they do: something to read, something to ask, something to call — or nothing at all.

Scope: the rubric measures businesses that sell products or services to buyers — firms whose customers ask questions, enquire, book or order. Media, government and reference sites sit outside its scope by design; their scores read as out-of-scope, not as verdicts.

The Ladder names five positions. A business sits on the highest rung it fully satisfies. Every rung is a position of opportunity: the point is never that you have failed a check, but that the next rung is specific, nearby, and nameable.

The five rungs

Rung	Name	You are here if…
0	Invisible	Agents are blocked in robots.txt, or your core content doesn't exist without JavaScript. Agents cannot reliably read you at all.
1	Readable	An agent can retrieve your pages and understand what you do.
2	Answerable	An agent can answer a buyer's real questions — what you offer, for whom, at what price band, with what availability — without a human stepping in.
3	Callable	At least one of your capabilities is invocable: an MCP endpoint, a documented API, or a registered WebMCP tool.
4	Transactable	An agent can complete a meaningful action end to end — book, order, submit, purchase — with human confirmation rather than human labour.

Rungs are gated: you cannot be Callable while Invisible — an endpoint nobody can discover doesn't count.

A note on rung 4: Transactable means an agent completes a real action end to end, and verifying that requires actually invoking a site's capability — which this scanner does not do uninvited. The automated scan therefore assesses up to rung 3; rung 4 is awarded through a consented invocation test, which is how attestation will work. Honest measurement beats a flattering guess.

What is captured

Five dimensions feed the rung. We publish what we observe — every scan lists its signals with the exact URL and timestamp behind each one. How those observations are weighted into scores is the rubric's method, refined against the benchmark corpus and versioned with every scan.

Dimension	What we capture
D1 Legibility	Can an agent read you at all? We observe robots.txt agent directives, llms.txt, content without JavaScript, sitemap, structured data, title/meta coherence — and whether you serve agents an optimised representation.
D2 Answerability	Can an agent answer a buyer's question about you? We observe price visibility, service definitions, FAQ coverage, and whether those pages are even discoverable.
D3 Callability	Is there anything an agent can invoke? We observe MCP endpoints (content-validated), WebMCP registrations, booking embeds — and forms, which are tools with a schema waiting to be written.
D4 Transactability	How far from intent to commitment? We observe contact affordances, friction markers, and the path to action.
D5 Standing	Would an agent trust and cite you? We observe entity clarity, named people, and consistency. (Deliberately coarse in v1.)

Two measurement rules distinguish this rubric from a checklist:

Evidence, not opinion. Every finding records what was observed, at which URL, at what time. A claim without a timestamp and a URL is not a finding.
Validated verdicts, not status codes. A URL that answers “200 OK” with a login page is not an MCP endpoint; an llms.txt that returns an HTML error page is not an llms.txt. Every probe checks what actually came back.

A worked example: rung 3 in the wild

Firecrawl's keyless launch removed the last human step between an agent and its product — no signup, no API key, no card. An agent can discover it, call it, and use it, this afternoon. That is what Callable looks like as a commercial decision rather than a technical aspiration — done by a real company, this quarter, as customer acquisition.

Versioning

The rubric is refined quarterly against the benchmark corpus. As the dataset grows, thresholds and weightings are re-cut to match how the agent-facing web actually behaves — a living standard, not a frozen checklist. Scores are comparable only within a rubric version: this is v1.0.0, changes are logged here, and every scan records the version it was scored against.

The first implementation of this rubric is the Agent Surface Scan — which is itself agent-callable, because a rubric about callability that an agent cannot call would be embarrassing.