Your Site's AI/Agent Readiness Score — and How We Built Ours

Classic SEO asks whether a page ranks. This asks something different: when an AI agent arrives — to answer a question, cite a source, or do something on a user’s behalf — how far does it get? That is a separate question from keyword rankings, and it is the one that decides whether your products show up in an assistant’s answer or get skipped because the agent could not read them. It is the data-and-structure side of the argument GEO for product data makes: being citable is overwhelmingly a data-quality and machine-readability problem, not a copy problem.

So we measured ourselves first, in public. Grade D is not a comfortable result to publish. It is also the honest one — most of the public web scores D or F on this rubric today — and publishing it is the point: the methodology below is reproducible, so you can run the same checks on your own site and get a number you can defend line by line.

The five categories and their weights

Each check returns a pass, partial, or fail, scored 0 to 1 (partials are proportional where a ratio exists). Checks that do not apply — Product schema on a site with no products, OpenAPI on a site with no API — drop out of the denominator so a site is not marked down for capabilities it has no reason to have. Category scores are weighted into an overall 0–100, and graded: A ≥ 90, B ≥ 75, C ≥ 60, D ≥ 45, F < 45. At 53, getclaro.ai is a mid-D.

What each category checks

AI Content Discovery (25%)

Everything downstream reads robots.txt first, so this category starts there and asks whether AI crawlers are actually allowed in.

• robots.txt present and parsable — a 404 is read by some crawlers as “do not crawl.”
• AI-bot directives: named groups for the AI crawlers, not just User-agent: *. The distinction that matters is retrieval vs training — OAI-SearchBot, Claude-SearchBot and PerplexityBot fetch to answer and cite now; GPTBot, ClaudeBot, CCBot and Google-Extended feed training. Block indiscriminately and you remove yourself from AI answers entirely.
• Content Signals (search / ai-input / ai-train) so you can allow citation while declining training.
• sitemap.xml so an agent enumerates you instead of guessing URLs.
• llms.txt with a described link per section — the descriptions are the point, because they let an agent pick the right page without fetching all of them.
• Meta robots not set to noindex.

Structured Data & Search Signals (20%)

• Valid JSON-LD (one malformed block can cause a consumer to discard the whole page’s structured data).
• Organization schema with name, url, logo, and sameAs — sameAs is how an agent reconciles the site with the LinkedIn, Crunchbase and GitHub entities and answers questions about the company.
• OpenGraph + Twitter cards for link-preview and summary text.
• FAQPage schema — the single highest-leverage addition for most B2B sites, because a question with a short declarative answer is exactly the unit an assistant quotes.
• Product / LocalBusiness / Article schema, scored only when commerce, physical-presence, or editorial signals exist.

Content & Semantics (20%)

• Server-side rendering: visible words in the initial HTML. Most crawlers do not run JavaScript, so a client-rendered site can look perfect in a browser and be an empty page to an agent. This is the most consequential — and most expensive — failure in the rubric.
• Heading hierarchy (one h1, no skipped levels — the tree is how a retriever chunks the page).
• Semantic HTML5, especially <main>: without it, an agent summarising the page can return your cookie banner.
• HTML lang, canonical URL, mobile viewport, image alt text, and freshness (lastmod / dateModified within ~60 days — a date that never changes reads as abandoned).
• Markdown content negotiation, which strips nav and script noise so more of an agent’s context budget lands on your content.

Agent Protocols & Discovery (25%)

This category is a forward bet, and almost every site on the public web fails most of it today — ours included. Everything else helps an agent read you; this lets it do something.

• MCP server card at /.well-known/mcp/server-card.json — the step change that lets an agent check availability, request a quote, or query a catalogue. Highest ceiling, highest cost.
• Link headers, agents.json / ai-plugin manifest, Agent Skills, API Catalog (RFC 9727), and WebMCP for in-page actions.
• OpenAPI and OAuth/OIDC discovery, scored only when an API surface is detected.

Security & Trust (10%)

• HTTPS with a valid certificate, HSTS, and a security.txt with a contact route. Small category, few checks — which, counter-intuitively, makes each check worth more points to recover.

Why we scored a D — honestly

A 53 is not one catastrophic failure; it is the ordinary shape of a site built for humans and search engines before agents were a consideration. The single biggest drag is Agent Protocols: at 25% of the weight and near-zero implementation across the whole web, it pulls almost everyone — us included — toward D or F. That is intentional. The rubric grades against where the web is going, not where its traffic is today, and a business that fails only on protocols is in a very different position from one whose pages are blank without JavaScript.

The rest of the gap is the familiar mix: discovery and structured-data signals that were never added because nothing used to read them. Which is exactly why we published the number instead of burying it — you can’t improve what you don’t measure, and a category breakdown tells you what to fix first far better than the headline grade does.

What we are fixing, and in what order

The methodology also ranks fixes by points recoverable against effort. The honest split:

This resources hub is one of the cheap wins already shipped: it serves an llms.txt with a described link per section, so an agent can pick the right page without fetching all of them. The structural items are on the roadmap, tracked against the same score so the next audit is a measurement, not a guess.

Run it on your own site

Point the five categories at your own homepage and site-level resources. For the page-level structure and signals an agent reads, the Product Page Report checks a product page directly, and three field references go deeper on the discovery layer: the AI crawler access checklist, the JavaScript visibility test, and the GA4 AI-visibility debug. The readiness of the site is one half; the readiness of the product records behind it is the other, which is the whole point of keeping one trusted product record that an agent can actually act on.

See how Claro keeps product records agent-readable

Related

FAQ

What is an AI/agent readiness score? +

It is a 0–100 measure of how well a website lets an AI agent discover, parse, cite, and act on it, scored across five weighted categories — AI Content Discovery, Structured Data & Search Signals, Content & Semantics, Agent Protocols & Discovery, and Security & Trust — and converted to a letter grade. It measures capability, not outcomes: whether a site makes citation and action possible, not whether an assistant currently does either.

What did Claro score on its own methodology? +

We ran getclaro.ai through our own methodology and scored 53 out of 100 — a Grade D. We are publishing that honestly because you cannot improve what you do not measure, and because most sites on the web today land in D or F territory for the same structural reasons, chiefly an Agent Protocols category that almost nobody implements yet.

Why is Agent Protocols weighted so heavily if almost no one passes it? +

Agent Protocols & Discovery carries 25% of the weight as a deliberate forecast of where the web is heading, not a measurement of today’s traffic. It is the category that drags most scores into D and F, including ours. A site that is strong everywhere else and fails only on protocols is in a very different position from one that fails on server-side rendering — the category breakdown matters more than the single number.

Is this a tool I can run? +

Today it is a reproducible checklist: the five categories and their checks are documented here so any team can run them by hand or script them. A hosted tool is planned, but the methodology is deliberately open so the score is auditable rather than a black box.