Google Open Knowledge Format (OKF): What It Is, the Spec, and How to Use It

Google published the Open Knowledge Format (OKF) on June 12, 2026, as version 0.1. It is a deliberately small, open format for writing down what an organization knows: its metrics, tables, datasets, APIs, and runbooks, in plain Markdown that any AI agent can read directly. Google calls v0.1 "a starting point, not a finished standard," and is clear that OKF is for AI agents, not a search ranking trick. This guide explains exactly what OKF is, the spec, what it is not, and how to decide whether to adopt it.

Last updated: August 2026

What OKF is, in one paragraph

OKF is a way to package an organization's knowledge as a bundle of Markdown files so that AI agents do not have to re-read and re-interpret the same internal documents over and over. Instead of pointing five different agents at the same wiki, the same spreadsheets, and the same API docs, you publish one shared "knowledge bundle" that every agent reads from. Each file in the bundle describes a single concept: one metric, one table, one API, one runbook. The format is intentionally minimal so it is easy to write by hand and easy for a model to parse.

The problem it solves

Most company knowledge lives in formats built for humans: dashboards, slide decks, Confluence pages, Notion docs, BI tools. When an AI agent needs a fact, like the exact definition of a metric or which API returns a given dataset, it has to crawl those human formats, guess at the structure, and spend tokens reconstructing meaning that a person already knows. That is slow, expensive, and error-prone, and every agent repeats the work. OKF flips it around: you state the knowledge once, in a machine-first format, and every agent reads the same canonical source.

The OKF spec: Markdown plus YAML frontmatter

An OKF bundle is a directory of Markdown files. Each file represents one concept and starts with a small block of YAML frontmatter, followed by free-form Markdown that explains the concept. In v0.1 the only required field is type. Everything else is optional.

Here is a complete OKF concept file. This one describes a metric:

---
type: Metric
title: Striking Distance Keywords
description: Keywords ranking 8-20 with quick-win potential.
resource: https://search.google.com/search-console
tags: [seo, content-audit, quick-wins]
timestamp: 2026-06-16
---
# Definition
Keywords with an average Google Search Console position between 8 and 20.
These are close enough to page one that small improvements can move them up.

That is the whole idea. A folder of files like this, one per concept, is an OKF bundle. Because it is just Markdown, you can keep it in a Git repository next to your code, review changes in pull requests, and let your agents read it the same way they read any other file.

What OKF is NOT

This is where most early confusion lives, so it is worth being blunt.

• It is not a search ranking signal. Adding OKF files will not improve your Google rankings. Google has said plainly that OKF is not a ranking or visibility mechanism, and there is no path from an OKF file to a blue-link boost.
• It is not a web-discovery standard. Unlike a sitemap or an llms.txt file, OKF is not something you publish at a well-known URL for the public web to crawl. It is designed as an internal knowledge bundle that your own agents read, not a public page format.
• It does not replace schema.org. Schema.org and JSON-LD structured data are still how you make web pages understandable to search engines and how you become eligible for rich results and Knowledge Panels. OKF sits in a different layer: internal agent knowledge, not public page markup.
• It is not finished. v0.1 is explicitly a draft. Expect the field set and conventions to evolve.

OKF vs llms.txt vs schema.org

These three are easy to mix up because they all involve making content readable to machines, but they target different jobs.

In practice they stack. Use schema.org so search engines and answer engines understand your pages. Use llms.txt so agents can navigate your public content efficiently. Use OKF so the agents you build internally stop re-deriving the same facts.

Who should actually adopt OKF

OKF earns its keep when you are building or running AI agents against your own data. Good candidates:

• Teams running internal copilots or RAG systems over company knowledge, where agents keep getting metric definitions or data sources slightly wrong.
• Data and analytics teams that want one authoritative definition of every metric, readable by both people and agents.
• Platform and API teams that want agents to discover and use their services without scraping documentation.

If you do not yet run agents against your own knowledge, OKF is a low-priority bet. The higher-leverage moves for being found and used by AI today are clean schema.org markup, a good llms.txt, and a real API or MCP endpoint.

How to get started in an afternoon

1. Pick five high-value concepts. Start with the things your agents get wrong most often: your top metrics, your core datasets, your main API.
2. Write one file per concept. Add the type field, a clear title and description, a resource link, and then a short Markdown body that states the definition precisely.
3. Keep the bundle in version control. Put the folder in your repo so changes are reviewed and dated. The optional timestamp field tells agents how fresh each fact is.
4. Point your agents at it. Load the bundle into your retrieval layer or give your agent file access to it, so it reads OKF before it guesses.
5. Expand from what breaks. Every time an agent gets a fact wrong, add or fix the concept file. The bundle grows to match where agents actually struggle.

Where this fits the bigger shift

OKF is one more sign that the web is being built for two audiences at once: people and the agents acting on their behalf. The same instinct shows up in agent-readiness signals like llms.txt, markdown content negotiation, MCP server cards, and machine-readable API catalogs. If you want to see how ready your own site is for AI agents, you can run a free agent-readiness scan, which checks for these signals and gives you copy-paste fixes.

Since v0.1 was published in June 2026, the developer community has responded quickly. Enterprise engineering teams and AI consultancies began publishing implementation guides within weeks, recommending phased pilots of 20 to 50 concepts with CI validation and freshness checks layered on top of the base spec. A productive debate has also emerged around OKF versus retrieval-augmented generation (RAG). The emerging practitioner view is that the two are complementary: OKF provides well-typed, explicitly described source files, giving your retrieval layer cleaner chunks to embed and reducing the misinterpretations that arise from vaguely labelled unstructured content. StartupHub.ai tracks over 71,000 startups and technology companies globally; across the AI agent tooling segment, the shift toward structured, agent-readable knowledge stores is one of the sharpest adoption patterns visible in 2026 infrastructure.

OKF v0.2: Trust Signals Added (July 2026)

Six weeks after v0.1 shipped, Google released OKF v0.2 on July 25, 2026. The update adds five opt-in fields designed to surface who produced a piece of knowledge, who verified it, and when it expires: provenance, trust, freshness, lifecycle, and attestation. As with v0.1, type remains the only required field; all new fields are optional and additive, so existing bundles are fully compatible without changes.

For most teams the v0.2 fields matter most for metrics and datasets where data quality and recency are critical. A financial model that uses stale revenue figures because an agent did not know the data was six months old is exactly the kind of silent failure the freshness and lifecycle fields are designed to catch. As of August 2026, no major AI agents are configured to read OKF bundles natively as a standard behavior, so adoption is still in early-practitioner territory. But the direction is clear: as agentic systems become responsible for more consequential decisions, provenance and trust metadata will matter as much as the facts themselves.

Frequently asked questions

Does Open Knowledge Format help SEO?

Not directly. OKF is not a ranking signal and is not crawled as public web content. For SEO and answer-engine visibility, use schema.org structured data and a good llms.txt. OKF helps the AI agents you run against your own data, not your Google rankings.

Is OKF the same as llms.txt?

No. llms.txt is a public file at your domain that tells AI crawlers how your site is organized. OKF is an internal Markdown bundle that your own agents read so they share one canonical source of knowledge. They complement each other.

What format does OKF use?

Markdown files with a small YAML frontmatter block. Each file is one concept. The only required field in v0.1 is type, with optional title, description, resource, tags, and timestamp.

Who created OKF and when?

Google published OKF v0.1 on June 12, 2026, within its Cloud ecosystem, describing it as a starting point rather than a finished standard. Version 0.2 followed on July 25, 2026, adding optional trust-signal fields: provenance, trust, freshness, lifecycle, and attestation. Existing v0.1 bundles are fully compatible and require no changes.

How does OKF compare to RAG (retrieval-augmented generation)?

OKF and RAG work together rather than against each other. RAG retrieves documents at query time to give models fresh context; OKF defines how those documents should be structured so retrieval is more precise. Well-typed OKF files give embedding models cleaner, consistently described chunks to index, which reduces the misinterpretations that come from vaguely labelled unstructured content. Most teams that adopt OKF use it alongside RAG: OKF standardizes the knowledge store, RAG surfaces the right files at runtime.

What changed in OKF v0.2?

OKF v0.2 (July 25, 2026) added five optional fields: provenance, trust, freshness, lifecycle, and attestation. These let you signal who produced a fact, how confident you are in it, how quickly it goes stale, and whether it has been reviewed. The type field is still the only required field. If you already have a v0.1 bundle, it works unchanged; v0.2 fields are additive.

Should my company adopt OKF now?

Adopt it if you are running AI agents or RAG systems against your own knowledge and they keep getting definitions or data sources wrong. If you are not yet building agents on your data, prioritize schema.org, llms.txt, and a real API first, and revisit OKF as the spec matures.