The four-group AI readiness model, explained

AIEO’s readiness score combines four groups of checks into one number. AIEO weights them 40/25/15/20 and publishes the reasoning, including the parts that are inconvenient for the score.

Scoring clarification, 3 October 2026: this article previously described WebMCP reference count as a scored check. It is informative. The explanation and example below now reflect the current scanner; the scoring rules have not changed with this correction.

Why a weighted score at all

Lighthouse solved this problem for page performance years ago: Total Blocking Time gets 30% of the score, Largest Contentful Paint 25%, Cumulative Layout Shift 25%, First Contentful Paint 10%, Speed Index 10%. Google says the weights were “chosen to provide a balanced representation of the user’s perception of performance,” with bands at 0–49 (red), 50–89 (orange) and 90–100 (green) — see Lighthouse’s performance scoring page.

AIEO borrows the shape, not the numbers: one overall figure built from named groups, each with a published weight, each auditable. An AIEO scan bands at 75+ (“Strong”), 50–74 (“Some gaps”) and below 50 (“Needs work”) overall, with a stricter 90/50 split for the individual groups — a group needs 90 to count as genuinely strong, not just passing.

The four questions our rubric asks

Our rubric covers discoverability, accessible content, structured data and static WebMCP source signals. These are AIEO’s chosen groups, not four prerequisites imposed by every AI system. Acting on a page is a separate capability from citing it:

The scan covers the homepage plus up to four discovered same-host pages (prioritising about, service, pricing, product, contact, solutions and what-we-do URLs), fetched with an honest AIEOBot/2.0 user agent — no browser disguise, one retry on network failure only — and robots.txt rules addressed to AIEOBot are honoured.

Why 40 for Traditional SEO

Google’s own AI-features guidance, updated 10 December 2025, is blunt: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” What it does recommend — crawlable robots.txt, internal links, textual content, structured data matching visible text — is standard technical SEO. Ranking and citation are different observations: Ahrefs found that 37.1% of cited URLs in its March 2026 sample appeared in the organic top ten for the same query. The often-quoted 37.9% counts the first ten SERP blocks, including features. The study’s parsing method also changed from the previous comparison. And the click itself is scarcer: Pew Research measured an 8% click-through to a traditional result when an AI summary was present, versus 15% without one. Those observations are reasons to measure both search traffic and AI visibility; they do not establish which matters more for every business.

AI Discoverability — 25 points, and llms.txt is unscored

This group scores search-lane bot access, the training-bot policy, crawl-delay, freshness, snippet controls, server-rendered content and an “answer-ready” paragraph directly after the first heading; llms.txt and pricing.md are reported alongside as informative. It’s tempting to assume llms.txt is the headline item here — some discussions of AI readiness give it considerable attention. AIEO doesn’t, on purpose.

Ahrefs audited 137,210 domains (15 June 2026) among sites using its analytics products and found 28% published an llms.txt file; 97% of those files received no requests during May 2026. GPTBot accounted for just 4.51% of the requests that did happen. Given that evidence, llms.txt and pricing.md are reported as informative and contribute zero points — not the 18–25 points some competing scorers hand the file. They’re worth knowing about, but the evidence doesn’t support treating them as pass/fail gates on the number itself.

What does carry real weight in this group is bot access. Cloudflare’s traffic data (29 August 2025) shows 79% of AI-bot crawling is for training, 17% for search, and only 3.2% is triggered by a live user action — with crawl-to-refer ratios as steep as 38,066:1 for Anthropic. The scan evaluates robots.txt under RFC 9309 per lane, with the limits of that vantage in mind. A Disallow on the search lane (OAI-SearchBot, Claude-SearchBot, PerplexityBot and their peers) is a fail under our rubric because it restricts a documented search-retrieval route; that finding is not a live citation test. A Disallow on the training lane (GPTBot, ClaudeBot, Google-Extended) is a warning. These policies have provider-specific effects: Google-Extended, for example, also governs certain Gemini grounding uses while leaving Google Search ranking unaffected.

Schema — 15 points, useful but not decisive

Schema checks whether JSON-LD is present, parses and carries a type, whether it includes an Organization or LocalBusiness block, and whether sameAs URLs are correctly formed. Its 15-point weight is our product choice, not an externally validated share of search value. Bing’s Fabrice Canel told Search Engine Land (20 March 2025) that schema markup helps Microsoft’s LLMs understand a page and that “gen AIs value fresh content… partly as a reference check.” Google’s guidance addresses eligibility for its own AI features: no special schema.org markup is required, and it curtailed FAQPage rich results to government and health sites in August 2023, then retired them for everyone from 7 May 2026 — the markup itself “can stay,” just don’t expect the rich result. AIEO’s v2 engine has retired scored FAQPage checks accordingly; FAQPage markup is now reported as informative, because rewarding it as if it still earned rich results would be misleading.

Agent Interaction — 20 points for static WebMCP evidence

This group has one scored check: WebMCP source evidence. It looks for a registerTool reference alongside modelContext/WebMCP context, or a form with toolname and tooldescription markup. With the other groups at 100, the outcomes are:

• No WebMCP context or declarative form: Agent 0, overall 80. An unrelated registerTool reference alone does not pass.
• API text without registration evidence: Agent 50, overall 90. The scan found a WebMCP/API mention, but no registration reference or declarative form.
• Qualifying registration or form source: Agent 100, overall 100. The scan found the source signature or declarative markup described above.

One qualifying reference and three qualifying references receive the same score. The reference count is informative, not a count of distinct or working tools: a loop can register several tools, while a library, comment or inactive code can contain unused references. JavaScript is not executed; browser support, runtime registration, discovery and successful tool calls remain unverified. Missing evidence within the scan limits does not prove WebMCP is absent or that other browser automation cannot use the site.

WebMCP is a W3C Web Machine Learning Community Group Draft Report, edited by engineers from Microsoft and Google; Chrome announced an origin trial in Chrome 149 on 9 June 2026. It is a draft, not a ratified standard. The 20% weighting is our product choice; neither that draft nor the origin trial establishes a measured contribution to business outcomes.

A worked example

Take a site with SEO 100, AI Discoverability 92, Schema 75, and no detected WebMCP source evidence (Agent 0):

40 (0.4 × 100) + 23 (0.25 × 92) + 11.25 (0.15 × 75) + 0 (0.2 × 0) = 74.25, rounding to 74 — “Some gaps,” just under the Strong band.

Add qualifying WebMCP source evidence and nothing else changes: Agent becomes 100, adding 20 points to the overall, taking it to 94 — “Strong.” Merely mentioning the API without registration evidence would score Agent 50 and give an overall score of 84. Adding more registration references after a pass earns no extra points.

SEO distributes its weight across multiple scored checks, while the Agent group currently scores just one. That large change reflects the chosen weights and thresholds, not demonstrated growth in visibility, enquiries or task completion. Our recorded enm-compass.life implementation moved from Agent 0 to Agent 100 after adding eleven tools. Eleven tools are not required for this score: that record illustrates source detection under the rubric, not runtime verification or an outcome benchmark.

What the score doesn’t know

Publishing the weights only means something if we also publish where they’re weak:

• Google-Extended is scored in the training tier even though it isn’t a crawler at all — Google’s own documentation says blocking it “does not impact a site’s inclusion in Google Search.” It also governs certain Gemini grounding uses, so this warning should be read as a provider-specific policy finding.
• User-fetcher rules are reported but unscored. OpenAI says robots.txt may not apply to ChatGPT-User and Perplexity says its user fetcher generally ignores it; Anthropic says Claude-User honours the file.
• Security headers (SPF, DMARC, HSTS, X-Content-Type-Options) are reported but carry zero weight in the score; they’re a posture badge, not a scored check.
• Everything is read from static HTML — there’s no headless rendering step. WebMCP detection follows same-site script imports up to six files deep, so a tool registered from a cross-origin bundle can be missed.

What to do

• Run a scan and read the group breakdown, not just the overall number. A low Agent score means qualifying WebMCP source evidence was not found or was incomplete within the scan limits. Neither group proves indexing, AI citations or successful use of the ordinary interface by an agent.
• If AI Discoverability is weak, check bot access first (robots.txt, Cloudflare bot rules) before worrying about llms.txt — access is scored, the file isn’t.
• Consider a WebMCP tool when it supports a useful customer task in your intended browser and agent. Test the whole journey; do not add tools just to raise the score.
• Don’t chase FAQPage schema for AI citations specifically — Google retired the rich result in May 2026, and AIEO reports it as informative rather than scored.

Run your own numbers at /scan and see which group is holding your score back.