The formal specification of the five metrics behind the AI Visibility Score
Purpose
This is the spec analysts and enterprise buyers cite when they want to know exactly what an AI Visibility Score measures. It defines the five metrics, what each one tests, why it matters to an AI answer engine, and how it rolls up into the 0 to 100 composite. It is deliberately implementation-neutral, it describes what is measured, not the proprietary scoring weights.
Core premise: AI answer engines don't reward pages, they reward legibility, the degree to which a machine can read a site, understand the entities on it, and cite them with confidence. Legibility is a measurable property. This framework is how you measure it.
The composite: AI Visibility Score (AIVS)
A single 0 to 100 number per site, rolled up from five weighted metrics. Deterministic, the same site scores within ±2 points on re-scan. Banded by the AI Visibility Maturity Model (Invisible / Partially Legible / AI-Legible / AI-Native).
The five metrics
1. Front Door, can a machine read the page at all?
Tests: whether AI and search crawlers can actually retrieve and read the content, not blocked by a bot-wall, not hidden behind JavaScript-only rendering, not served an error or empty page with a success status.
Why it matters: it is the precondition for everything else. A site that fails the Front Door is invisible regardless of how good its content or schema is. It is binary and it is first.
Plain-English name: The Front Door Test.
2. Entity Integrity, is the brand and each product a clean, machine-readable fact?
Tests: the completeness and validity of structured data, Product (price, availability, ratings, identifiers), Organization (name, verified identity links), and their consistency with the visible page.
Why it matters: AI cites entities, not prose. If the entity layer is thin, broken, or contradicts the page, the model has nothing reliable to quote, or worse, quotes something wrong.
Sub-metric: Entity Density, how completely each entity is described.
3. AI Readability, is the answer stated, or buried?
Tests: whether key facts are expressed in an extractable form, clear definitions, direct statements, structured specs, rather than buried in marketing prose.
Why it matters: answer engines extract; they don't infer through ambiguity. Content that states the answer plainly gets used; content that hides it gets skipped.
Plain-English name: Reading Level.
4. Mesh Integrity, do the pages form one coherent map?
Tests: whether collections, products, and key pages are connected into a navigable graph that reveals how things relate, not a pile of disconnected islands.
Why it matters: the mesh is how a model understands "this brand sells X, for use case Y, related to Z." A broken mesh means the model sees fragments, not a business.
Plain-English name: The Mesh.
5. Authority Signal, is the identity consistent, current, and trustworthy?
Tests: freshness, identity consistency across the site and its structured data, and the presence of trust markers a model uses to decide how confidently to cite a source.
Why it matters: legibility gets you read; authority gets you trusted enough to be named.
Plain-English name: Freshness.
How they combine
AI Visibility Score (0 to 100)
│
├── Front Door ── precondition (gates the rest)
├── Entity Integrity ── the facts AI cites
├── AI Readability ── extractability of those facts
├── Mesh Integrity ── the relationships between them
└── Authority Signal ── confidence to cite
A site scores well only when the foundation (Front Door) is open AND the entity layer is clean AND it's all extractable, connected, and trustworthy. Weakness in any single metric caps real-world AI visibility, which is why the field's scores are bunched: almost every site is strong somewhere and broken somewhere.
What the framework deliberately does not publish
The exact scoring weights, breakpoints, and sub-criteria are proprietary (they are what make the score reproducible and defensible). This spec defines the measured concepts so the framework is transparent and citable without giving away the methodology, the same way Core Web Vitals defines LCP/CLS/INP publicly while Google's exact thresholds evolve.
Versioning
This is v1.0. The framework is versioned so research published against it is always reproducible: a score is only comparable to another score under the same framework version. Changes ship as v1.1, v2.0, etc., with a changelog.
Cite as: Visibility Mesh, AI Legibility Framework v1.0 (2026).