# Agent Readability Spec ARS

[4 · Read](/category/read/)

| Agent Readability Spec |
| --- |
| Category | [Read](/category/read/) |
| Governed by | Vercel; published in Vercel's own Knowledge Base with no external standards process |
| Status | A specification and scoring checklist, not a protocol; framed as best practices to implement voluntarily |
| Phase | draft |
| Implementation | shipping |
| Adoption | unmeasured |
| Version | current (last updated 2026-03-23 per a secondary source; no version number or publication date is given on the page itself) (2026-03) |
| Specification | [https://vercel.com/kb/guide/agent-readability-spec](https://vercel.com/kb/guide/agent-readability-spec) |
| Found at | none |
| Acronym collisions | Agent-Friendly Documentation Spec (a separate, earlier spec by a different maintainer covering similar ground — see dispute) |
| Last checked | 2026-09-06 |

[Read](/category/read/) Vercel's Agent Readability Spec is a checklist and 0-100 scoring system for how well a website's content can be discovered, fetched, and understood by AI agents [[1]](#ref-1). It groups roughly twenty independently verifiable checks into three areas: discovery (an llms.txt file, a permissive robots.txt, XML and markdown sitemaps), structure (canonical links, meta tags, Schema.org/JSON-LD, heading hierarchy, text-to-HTML ratio), and context (an AGENTS.md skill file, markdown mirrors of HTML pages, Accept: text/markdown content negotiation, fenced and language-tagged code blocks, OpenAPI/Swagger schema links) [[1]](#ref-1).

## Overview
It solves the problem of turning "be agent-friendly" into a concrete, curl-verifiable punch list: each requirement in the spec ships with an example and a shell command to check it, and site owners get a single percentage score (passed / total × 100) they can track over time [[1]](#ref-1). Notably, it documents — by name, as Vercel's own recommended practice — serving a markdown alternate for every HTML page and returning it when a client sends Accept: text/markdown, with a Link header pointing back to the canonical HTML URL [[1]](#ref-1). This is the same content-negotiation pattern also described in llms.txt v2's discovery mechanism (see the llms-txt record) and is the clearest documented instance of the general "markdown twin" practice found during this verification pass.

It does not solve trust or accuracy: nothing in the spec checks whether the content is correct, and a perfect score says only that a page is structurally easy to fetch and parse, not that an agent's answer using it will be right. It is also not a protocol or a machine-readable manifest — there is no fixed path an agent fetches to learn a site's score; the checklist is meant to be run by (or on behalf of) the site owner, and scoring is reproduced by third-party tools rather than looked up from a canonical registry.

Current state, as of this verification (September 2026): the page itself carries no publication date or version number. An independent commentator dated it to "last updated March 23" (2026) and noted its author is Timothy Jordan [[2]](#ref-2). At least one third-party checker, agent-ready.dev, implements scoring explicitly described as checking a site "against the Vercel Agent Readability Spec and llmstxt.org standard," which is the clearest evidence of independent implementation found [[3]](#ref-3); Vercel's own knowledge-base article is the only primary specification text.

What is disputed or unknown: the spec was published unilaterally by a hosting vendor, with no external review process, and overlaps substantially with an earlier, separately maintained 'Agent-Friendly Documentation Spec' — a secondary source flagged this as worth scrutinizing given Vercel's commercial interest in the hosting practices the spec implicitly favors (server-rendered pages, content negotiation infrastructure) [[2]](#ref-2). No independent adoption count or measurement of how many sites meet a given score threshold was found.

## Discovery and probe

none — not a discoverable manifest; it is a checklist a site owner applies to their own pages, verified externally by tools like agent-ready.dev or is-agentic.com · none

```
GET {origin}/ (plus /llms.txt, /robots.txt, /sitemap.xml, /sitemap.md, /AGENTS.md as sub-checks; see the spec's own per-requirement curl commands)
parse:     No single field proves conformance; the spec defines ~20 discrete site- and page-level checks (llms.txt served, robots.txt allows AI bots, AGENTS.md present, canonical link, JSON-LD, markdown mirrors, Accept: text/markdown negotiation, etc.) each independently verifiable by curl, and a numeric score = round(passed/total * 100)
on absent: A site failing most checks is not in violation of anything; the spec is a voluntary readiness checklist, not an enforced requirement
```

Not probed live: no-fixed-path. A checklist, not a manifest; the llms-txt, agents-md, robots and sitemap checks cover its site-level items.

## History
- 2026-09-06 — new record added during read-step verification; verified against vercel.com/kb/guide/agent-readability-spec and independent commentary

## Instances

- [is-agentic.com (separate agent-readiness scoring API referenced in Vercel's own follow-up guide)](https://vercel.com/kb/guide/make-your-site-readable-by-ai-agents) · Recorded 2026-09-06 · service
- [agenticweb.wiki](https://agenticweb.wiki/llms.txt) · Observed 2026-09-06 · site
- [agent-ready.dev (scores sites against the Vercel Agent Readability Spec and llmstxt.org)](https://agent-ready.dev/) · Recorded 2026-09-06 · service
- [agent-ready-mcp](https://github.com/mlava/agent-ready-mcp) · Observed 2026-05-11 · service
- [a14y (independent implementation)](https://github.com/jesserobbins/a14y) · Observed 2026-05-09 · service

Every instance with its source is listed on the [instances page](/instances/).

## See also

- Not to be confused with: Agent-Friendly Documentation Spec (a separate, earlier spec by a different maintainer covering similar ground — see dispute)

## References

- Full checklist text: site-level files (llms.txt, robots.txt, sitemap.xml/md, AGENTS.md), page-level HTML requirements (canonical link, meta tags, JSON-LD, heading structure, signal-to-noise ratio), server configuration (markdown mirrors, `Accept: text/markdown` content negotiation, `Link` header canonical), content quality (fenced code blocks, OpenAPI schema links), and the scoring formula — [https://vercel.com/kb/guide/agent-readability-spec](https://vercel.com/kb/guide/agent-readability-spec) (2026-03) VERIFIED
- Independent commentary identifies the author as Timothy Jordan, notes the page carries no publication date and was 'last updated March 23' (2026), and frames it as a competing effort alongside a pre-existing 'Agent-Friendly Documentation Spec' maintained separately — [https://dacharycarey.com/2026/03/28/how-to-evaluate-platform-written-spec/](https://dacharycarey.com/2026/03/28/how-to-evaluate-platform-written-spec/) (2026-03-28) REPORTED
- A third-party checker, agent-ready.dev, scores sites specifically 'against the Vercel Agent Readability Spec and llmstxt.org standard', indicating at least one independent implementation of the scoring logic — [https://agent-ready.dev/](https://agent-ready.dev/) (2026) REPORTED

Disputed: Vercel published this spec without an external standards process or a clear version/date, prompting at least one independent commentator to note it overlaps with, and was not coordinated with, a pre-existing 'Agent-Friendly Documentation Spec' maintained by someone else.

[JSON](/data/protocols/vercel-agent-readability.json) · [Markdown](/protocols/vercel-agent-readability/index.md)