What Core Agent Vitals measures
Google's Core Web Vitals measure how a page performs for a human eye — how fast it paints, how stable it stays as it loads. But AI agents now read the web for your users: to answer a question, compare options, summarise a page, or take an action on their behalf. Core Web Vitals say nothing about that reader. A page can score a perfect LCP for a person and be nearly useless to an agent — its content injected by JavaScript a crawler never runs, its entities undeclared, its meaning buried under a token budget of navigation and boilerplate.
Core Agent Vitals (CAV) is the missing layer: six metrics for how well an AI agent can find, recover, understand, and act on your page. The analyzer above measures every one of them against any URL. Here is what each metric means, and why it decides whether an agent can actually use your site.
The six metrics
CRR — Content Recovery Ratio
How much of your page's meaningful content an agent recovers from the raw HTML, before any JavaScript runs, versus after a full render. Many AI crawlers do not execute JavaScript at all — if your content is injected client-side, they see an empty shell. A client-rendered page can look flawless to a person and score near zero for an agent. This is the most direct measure of whether an agent can even read you. Good is 0.95 or higher.
SSD — Semantic Signal Density
How much of your page is real content versus navigation, headers, footers, and boilerplate, combined with how complete your structured data is. An agent pays for every token it reads; a page that is mostly chrome wastes its budget before it reaches your content, and undeclared entities force it to guess who and what the page is about. Dense, well-marked-up content resolves cleanly. Good is 0.60 or higher.
ARR — Action Resolution Rate
The share of your critical actions — search, add to cart, subscribe, book — that an agent can resolve through stable, accessible locators with zero selector drift. Your accessibility tree is the agent's API. If a button is reachable only through a brittle CSS path that changes between deploys, an agent cannot reliably act on it, even when a human can. Good is a perfect 1.0.
TC — Token Cost
The number of tokens it costs an agent to read your page's content. A bloated, boilerplate-heavy page burns the agent's context window before it reaches what matters, and costs more money on every single request an agent makes against it. Lean, content-first pages are cheaper and faster for a machine to consume. Good is under 4,000 tokens.
TTFUT — Time to First Useful Token
How quickly the first useful piece of content reaches a streaming agent. Slow server responses and render-blocking resources delay when an agent can begin working with your page — the machine equivalent of a slow first paint. It is a latency signal to watch rather than a hard gate, because it varies with network conditions.
AF — Answer Fidelity
The north-star metric: given only your page's agent representation, can a model correctly answer canonical questions about it? This measures whether your page is not just recoverable but genuinely understood — the difference between an agent seeing your text and an agent getting your meaning right. A page can be perfectly recoverable and still be misread. Good is 0.95 or higher.
How the score works
Your overall score weights these metrics along the agent's journey — find, recover, understand, act — into a single number from 0 to 100. On top of the weighting sit deterministic gates for the failures that invalidate everything else: a page that serves agents materially different content than it serves browsers (cloaking), blocks AI crawlers outright, or returns only a thin shell to a scanner can never reach a passing score, regardless of how good its other numbers look. The gates keep the score honest — a site cannot look agent-ready while quietly being closed to agents.
Why it matters
An agent that cannot recover your content will describe your business from a guess — often from your page title or navigation menu, not your actual offering. An agent blocked at the door cannot cite you at all. As more discovery, comparison, and buying moves through AI assistants, the sites an agent can read, understand, and trust are the ones it surfaces and recommends. The sites it cannot are simply left out of the answer.
A metric is only worth watching if it predicts something real. The cheapest one — CRR — reliably predicts whether a model can recover facts from a page: across 46 pages it separates the readable from the invisible at ROC AUC = 0.95, with synthetic canaries confirming the model reads the page rather than reciting what it already knew (priors-leak 0.00). See the full spec and method.