How We Scored 100/100 on Agent-Readiness (2026)

TL;DR: ModelFit scored a perfect 100/100, Level 5 "Agent-Native" on isitagentready.com, up from 27 before we started. We passed 14/14 checks across four scored categories by shipping the emerging agent-web standards: llms.txt, Markdown content negotiation, Content Signals, an RFC 9727 API catalog, an OAuth2 client-credentials server, a live MCP endpoint, and a signed Agent Skills index. This post is the exact build list — and an honest take on what the score does and does not buy you.

Most websites are built for human eyes and Googlebot. A growing slice of traffic is neither — it is an AI agent fetching your page on a user's behalf, reading it, and sometimes acting on it. ModelFit's own logs show ChatGPT-User crawling us roughly 30 times more often than Googlebot. So we spent two weeks making the site readable, callable, and trustworthy to agents, then measured the result on a public scanner. The score went from 27 to 100. Here is what we shipped and why each piece matters.

What does an "agent-readiness" score measure?

It measures whether an AI agent can discover, read, access, and authenticate against your site using emerging open standards. The isitagentready.com scanner buckets its checks into five categories: Discoverability, Content, Bot Access Control, "API, Auth, MCP & Skill Discovery," and Commerce. Each is a yes/no audit of a specific file, header, or endpoint — not a quality judgment of your prose.

That distinction matters. The score rewards plumbing: can a machine find your llms.txt, negotiate Markdown, read your robots directives, and discover your tools? It says nothing about your rankings, your traffic, or whether anyone actually calls those endpoints yet. We hit 100 on four categories and skipped Commerce (it needs agentic checkout infrastructure we do not sell through).

How do agents discover a site? (Discoverability: 4/4)

Agents discover capabilities the way browsers discover stylesheets — through advertised links and well-known files. We shipped four discovery surfaces.

• llms.txt at the root. A plain-text index of our important URLs with descriptions, kept under 10KB. It is the agent equivalent of a sitemap written for a reader, not a crawler.
• RFC 8288 Link headers on the homepage: describedby points to llms.txt, service-doc to our about page, and api-catalog to the well-known catalog. An agent reading response headers learns where everything lives without parsing HTML.
• An RFC 9727 API catalog at /.well-known/api-catalog, listing our OpenAPI service description and a health-status endpoint.
• Standard sitemap and metadata signals so traditional crawlers stay happy too.

The principle: never make an agent guess. Every capability has a machine-findable pointer.

Why serve Markdown instead of HTML? (Content: 1/1)

Because HTML is expensive for an agent to read, and Markdown is not. We added content negotiation: when a request carries Accept: text/markdown, our middleware returns a clean Markdown version of the homepage and any blog post instead of the full HTML page.

The payoff is tokens. Stripped of navigation, scripts, and styling, a Markdown page costs an agent a fraction of the tokens to ingest — we return an x-markdown-tokens header so the client knows the cost up front. Same content, lower bill, faster reasoning. For a site that wants to be cited by ChatGPT and Perplexity, cheap-to-read pages are a direct advantage.

What is Bot Access Control? (Bot Access Control: 2/2)

It is telling agents not just whether they can crawl, but what they may do with what they read. Two pieces.

First, a proper robots.txt that allows the AI crawlers — GPTBot, ClaudeBot, PerplexityBot and friends — rather than the default-deny many CDNs apply. Second, and newer, Content Signals: a directive on our robots route that emits Content-Signal: search=yes, ai-input=yes, ai-train=no. Translation: yes, use our pages to answer live questions; no, do not train base models on them. It is a granular consent layer the older allow/deny model could not express.

How do agents call your tools? (API, Auth, MCP & Skills: 7/7)

This is the deep end — where a site stops being a document and becomes a service an agent can invoke. It was also the bulk of the work, and where our score moved the most. Seven checks, all green.

1. OpenAPI 3.1 service description — a machine-readable contract for our API, linked from the catalog.

2. OAuth2 client-credentials server — agents register, get a token, and authenticate. Tokens are deliberately narrow: they grant only an agent:identify scope and cannot touch private data.

3. A protected resource endpoint that proves the auth flow works end to end.

4. An agent_auth discovery block in our well-known metadata, pointing at a human- and machine-readable auth.md.

5. A live MCP server at /api/mcp/ — a real Streamable-HTTP JSON-RPC endpoint implementing the Model Context Protocol. It exposes three tools an agent can call: recommend_local_models, list_supported_hardware, and get_status.

6. A signed Agent Skills index — a .well-known/agent-skills/index.json listing skills (like find-local-llm) with SHA-256 digests so an agent can verify it fetched the real, untampered skill.

7. WebMCP — the same tools registered in-browser via navigator.modelContext, so an agent driving a real browser session can call them directly.

The headline here is the MCP server. An agent that speaks MCP can ask ModelFit "recommend a local model for a 16GB M2 Mac" and get a structured answer back — no scraping, no HTML parsing, just a tool call. That is the difference between being a page an agent reads and a service an agent uses.

Does a 100 score boost SEO or traffic?

No — and pretending otherwise would be dishonest. Agent-readiness and search ranking are different systems. Google ranks on relevance and E-E-A-T; the agent-readiness scan never looks at those. A perfect 100 will not move you up the SERP and will not, by itself, add a single human visitor.

What it does is more like laying fiber before the neighborhood fills up:

• Near term: cleaner, cheaper-to-read pages mean marginally more and better AI citations in ChatGPT, Perplexity, and AI Overviews. Measurable, but small.
• Medium term (6–18 months): as agentic browsing and tool-calling grow, the infrastructure is already built. If MCP-driven discovery or agentic commerce becomes mainstream, we are positioned, not scrambling.
• Today's real ROI: the milestone itself — a public, verifiable proof point that ModelFit is built for the agent web.

It is a moat dug early, not a faucet turned on. We think that is the right bet for a site whose entire audience is people running local AI.

How can you check and improve your own site?

Start by scanning, then close gaps category by category. The practical path:

1. Run your URL through a public agent-readiness scanner to get a baseline (ours was 27).

2. Discoverability first — add an llms.txt and Link headers. Cheap, high-signal, one afternoon.

3. Content — add Accept: text/markdown negotiation so agents can read you cheaply.

4. Bot Access Control — fix robots.txt to allow AI crawlers, then add Content Signals for consent granularity.

5. API/Auth/MCP — the heavy lift. Publish OpenAPI, stand up an MCP endpoint for any real capability you offer, add OAuth if agents need authenticated access, and sign your Agent Skills.

Most sites can reach a respectable score on the first three categories quickly. The fourth separates documents from services — only worth it if you have a tool or API genuinely worth exposing.

FAQ

What does "Level 5 Agent-Native" mean?

It is the top tier on isitagentready.com's scale, awarded when a site passes the full battery of discovery, content, access-control, and API/auth/MCP checks. It signals the site is built to be used by agents, not merely crawled by them.

Will this help me rank higher on Google?

No. Agent-readiness is independent of search ranking. Google ranks on relevance and E-E-A-T signals the scan does not measure. Treat agent-readiness as preparation for AI-agent traffic and citations, not as an SEO tactic.

Do I need an MCP server to get a good score?

Not for a good score — Discoverability, Content, and Bot Access Control get you most of the way. You need MCP, OAuth, and an Agent Skills index to max the "API, Auth, MCP & Skills" category and reach Level 5. Only build those if you have a real capability worth exposing to agents.

What is the difference between an agent reading my site and using it?

Reading means ingesting your HTML or Markdown to answer a question — a citation. Using means calling a tool you expose (via MCP or an API) to perform an action or fetch a structured answer. A page can be read; a service can be used. We built both.

Is the score worth chasing if no agents call my endpoints yet?

It depends on your audience. For most businesses, the first three categories are worth it now for cleaner AI citations. The full Level 5 build pays off only as agent adoption rises — it is a forward bet, valuable if your users are early on the agent curve, as ours are.