The Open Knowledge Format (OKF), explained

What is the Open Knowledge Format?

The Open Knowledge Format is an open, human- and agent-friendly format for representing knowledge: the metadata, context, and curated insight that surround data and systems. Google Cloud's Data Cloud team, led by Sam McVeety and Amir Hormati, published OKF v0.1 in June 2026 in the public Knowledge Catalog repository. It formalizes the LLM-wiki pattern that Andrej Karpathy described (a markdown knowledge base an agent reads and maintains like code) into a portable, vendor-neutral standard.

The point is producer and consumer independence. A bundle hand-written by a person can be read by an AI agent. A bundle generated by one tool can be browsed in another. Because it is just markdown, just files, and just YAML frontmatter, anyone who can open a file can read it and anyone who can clone a git repo can ship it. OKF is the file format agents read, not a service you call.

Where OKF fits: the agent-readable web stack

OKF is the top floor of a stack of conventions for making a site legible to machines. Each layer hands an agent a little more, and the lower layers only pay off when there is substance up here: Ahrefs found llms.txt adoption grew 8.8x in the year to mid-2026 while 97 percent of the files received zero AI fetches. A pointer with nothing behind it is just a pointer.

A sitemap says which pages exist. llms.txt and AGENTS.md or CLAUDE.md point and instruct. OKF goes one floor higher and hands over the content itself, as a graph an agent can traverse. For the full picture of making a site legible to AI, see our guide to AI search optimization.

Why Google built OKF

In most organizations, the knowledge an AI agent needs lives in fragmented surfaces: metadata catalogs with their own APIs, wikis and shared drives, code comments and docstrings, and the heads of a few senior engineers. When an agent has to answer a question like how weekly active users are computed from the event stream, it reassembles the answer from scattered, mutually incompatible places, every time.

Google's argument is that the fix is not another knowledge service. It is a format: a way to represent knowledge that anyone can produce without an SDK, anyone can consume without an integration, that survives moving between systems, lives in version control next to the code it describes, and is readable by humans and parseable by agents in the same file. That format is OKF.

The OKF format in depth

A bundle is a directory tree of markdown files. This site ships a real one in its repository, 26 concepts joined by 128 links across 10 concept types, and trimmed to its folders it looks like this.

Each concept is one .md file. The concept id is its bundle-relative path, so the file's location is its identity, and links written in the absolute bundle-relative form (/services/technical-seo.md) survive file moves, which is the form spec v0.2 recommends. A YAML frontmatter block carries the structured fields, and the body is normal markdown.

---
type: Service
title: AI search optimization
description: Get a brand cited by ChatGPT, Perplexity, and Google AI Overviews.
resource: https://witscode.com/ai-search-optimization
tags: [seo, aeo, ai-search]
status: stable
generated:
  by: human:sudhakar
  at: 2026-08-12T09:00:00Z
---

# What it is

Structuring content and entities so large language models cite you.

# Related

Pairs with [technical SEO](/services/technical-seo.md) for crawlability.

The only required field is type, a short non-empty string naming the kind of concept. Two filenames are reserved: index.md lists a folder's concepts for progressive disclosure, and log.md records dated changes.

# Knowledge bundle

## Services
* [Web application development](/services/web-application-development.md) - custom web apps
* [AI search optimization](/services/ai-search-optimization.md) - get cited by AI

## Pipelines
* [Lead pipeline](/pipelines/lead-pipeline.md) - how website leads reach the CRM

Internal links are the important part. A link from one concept to another is a directed edge, so the directory becomes a graph. An agent does not just learn what each concept is, it learns how they sit relative to each other, which is most of what understanding a system actually means.

OKF field reference

The full set of frontmatter fields. Only type is required. Everything else is recommended or optional, and producers may add their own keys.

How agents read an OKF bundle

OKF separates two roles. Enrichment agents write into a bundle, drafting concepts from a database, a codebase, or a catalog and adding citations and schemas. Consumption agents read and traverse it. Because the contract between them is just files, neither has to know anything about the other.

A consumption agent typically starts at the root index.md, reads the entries it needs, and follows links deeper, a pattern Google calls progressive disclosure. This is the same instinct behind a shared markdown library that agents read and update over time, which is where production AI agents and MCP servers are heading: less re-deriving the same facts, more reading a curated source of truth.

How to add OKF to your website and WordPress

You add OKF by publishing a bundle alongside your site. There are three common paths.

On WordPress

WordPress has no official OKF integration yet, so the practical route is a generator: map your post types and pages to concept types, emit one markdown file per concept with frontmatter, serve the bundle at /okf/, and rebuild on publish so it never goes stale. The generation is the easy, mechanical part. The value is in typed concepts (a Service, a Metric, a Runbook) linked to each other the way they actually relate, which no automatic export can decide for you. The same is true whether you run WordPress, headless, or a custom web app.

OKF vs llms.txt, AGENTS.md, and CLAUDE.md

These conventions solve different problems. llms.txt is a pointer, AGENTS.md and CLAUDE.md are instructions, and OKF is the portable knowledge itself.

Is OKF worth it? An honest take

First, clear up the common myth: OKF is not an SEO ranking signal. Google's search systems do not fetch a bundle from your site and rank you on it. It is an internal knowledge format for agents, not a web publishing signal. A bundle will not move your rankings this week, or next week.

It is also early. Google still frames the spec, now at v0.2, as a starting point, not a finished standard. The spec is deliberately lenient (broken links and unknown types are tolerated), which buys resilience as a wiki grows but means quality varies a lot between bundles. And an agent-updated knowledge base is a real attack surface: if an agent writes from untrusted input, a bundle becomes a vector for indirect prompt injection, so what you let write into it matters.

The honest case for adopting now is the schema-markup case from a decade ago. It is cheap to ship, it makes your knowledge legible to the agents and answer engines that are starting to field questions about you, and shipping early is how you learn the format before it matters. A small, well-typed, validated bundle beats a large auto-generated one. If you ship anything, ship that.

Latest OKF updates

OKF is a moving spec, already at v0.2. We keep this guide current as it evolves. The notable milestones so far:

Validate your OKF bundle

A bundle is only useful if any tool can read it, so check it against the spec before you publish. WitsCode maintains a free, open-source conformance suite and validator that tells you, pass or fail, whether a bundle follows the Open Knowledge Format. It tracks spec v0.2 and is verified against Google's own reference bundles. The suite also ships a graph tool (okf-graph.mjs) that renders any bundle as an interactive visualization in a single self-contained HTML file, and an agent skill that teaches a coding agent to produce, consume, and maintain a bundle without drifting from the spec.

Open Knowledge Format FAQ