State of the Agentic Web

Access, understand, act — across the dataset

The same deterministic regrouping every individual report uses, aggregated across every eligible website's latest scan. Median points shown, not mean — a small dataset is more easily skewed by a handful of extreme scans.

• Access15 / 15
• Understand24 / 40
• Act1.5 / 45

How much of the observed web is agent-actionable?

Averaged across the same 15-capability vocabulary every scan checks: how often a capability is available to a human, described in a machine-readable way, and directly callable by an agent. These are three independent measurements, not a funnel — a capability can be machine-readable without being agent-callable, or vice versa.

• 11%Human-accessible
• 1%Machine-understandable
• 3%Agent-actionable

Where the biggest human-vs-agent gaps are

Capabilities most often available to a human visitor but not (yet, or verifiably) callable by an agent — the human-detected rate minus the agent-callable rate, across the eligible dataset. This does not imply causation between capabilities; each is measured independently.

• Contact74% human→0% agent
• Search48% human→13% agent
• Compare13% human→0% agent
• Availability9% human→0% agent
• Get quote4% human→0% agent

Full 15-capability human/machine/agent breakdown is available on every individual report.

WebMCP adoption

WebMCP is early: a W3C Community Group proposal with one shipping implementation (Chrome, from Chrome 146 Canary, February 2026, behind an origin trial). No other major browser ships it yet, and there is no centralized directory of which sites implement it.

• Detected9.4%
• Unable to verify81.3%
• Not detected9.4%

Why detection is hard for a static scanner

WebMCP tools are registered by JavaScript at runtime (navigator.modelContext.registerTool()). A scanner that only fetches and reads HTML — which is what MCPWebsite's scanner does, deliberately avoiding headless-browser automation for cost and scope reasons — cannot execute that code. It can only look for the literal API call in the scripts it already has, or in an external script the page itself references.

We never collapse the three resulting states into a flat "no":

• Detected — a literal registerTool() call was found and could be parsed, either inline or via a referenced external script.
• Unable to verify — the page references WebMCP-adjacent code we couldn't fully inspect, so we can't rule it in or out.
• Not detected — no related signal was found anywhere we could read.

Distribution of Agent Readiness scores

How the eligible dataset spreads across the 0–100 score range — the shape behind the median above.

• 0-19
  4 (13%)
• 20-39
  11 (34%)
• 40-59
  14 (44%)
• 60-79
  1 (3%)
• 80-100
  2 (6%)

Methodology

Eligibility. A scan counts toward every number on this page when it completed (fully or partially) against real public-scanner traffic — internal, cohort, and future API-driven scans are excluded, so this page never reflects MCPWebsite's own test activity.

Latest scan per website. A website with multiple scans contributes only its most recent one, filtered to the current scoring version (v1.0.0). A website whose latest scan predates a scoring-version bump is excluded entirely until it's rescanned, rather than being represented by a stale, incomparable score.

Access / Understand / Act. A deterministic regrouping of the same five scoring categories every report shows — Access is AI Accessibility; Understand is Machine Understanding plus Business Information; Act is Agent Actions plus WebMCP. This is presentation only; it never changes the underlying score.

Deterministic evidence vs. AI interpretation. Every number on this page — scores, capability statuses, WebMCP status — comes from deterministic, rule-based scanning (scanner v1.0.0, scoring v1.0.0). AI (analysis v1.1.0) is used only on individual unlocked reports to interpret and prioritize that evidence in plain language — it never determines a score and never appears in the aggregate numbers on this page.

Limitations. The dataset is small and skews toward whichever websites visitors have chosen to scan — it is not a random or representative sample of the web, and detailed breakdowns are withheld below 20 eligible websites for exactly that reason. Treat every number here as directional until the dataset is much larger.

As the specification stabilizes and adoption grows, the numbers above will keep tracking real scan data — no synthetic or projected figures.