The Open Knowledge Format: What OKF Means for Modern Search and Content

What exactly is the Open Knowledge Format?

OKF formalises a pattern that emerged organically wherever people started feeding knowledge to language models: a folder of markdown files, one file per concept, each with a small YAML metadata block on top. The specification, published by Google Cloud in its knowledge-catalog repository, is deliberately tiny. A knowledge bundle is a directory tree. Each concept file needs exactly one required field, type, plus recommended title, description, resource (the canonical URI of the thing described) and tags. Reserved index.md files give agents progressive disclosure, a listing of what exists before opening anything, and log.md files record change history. Concepts cross-link with ordinary markdown links, which turns the bundle into a knowledge graph. There is no schema registry, no central authority and no required tooling; the spec's own summary is that if you can read a file you can read OKF, and if you can clone a repository you can ship it.

It was built for a specific pain Google calls the context assembly problem: organisational knowledge, table schemas, metric definitions, playbooks, policies, is scattered across catalogs, wikis, code comments and people's heads, and every team building an AI agent re-solves the job of gathering it. OKF gives producers (people, agents, export pipelines) one portable format to write into, and consumers (agents, search indexes, interfaces) one format to read. That origin matters for how the SEO industry should interpret it, which we come to below.

---
type: Article
title: The Perfect HTML Page
description: How markup decides which page search engines, AI engines and humans choose.
resource: https://www.laurelinlabs.com/blog/perfect-html-page
tags: [on-page-seo, aeo, page-structure]
generated: { by: laurelin_agent/model-id, at: 2026-08-23T20:00:00Z }
verified: { by: human:piers, at: 2026-08-23T20:30:00Z }
stale_after: 2027-02-23T00:00:00Z
---
The distilled knowledge of the page, in structural markdown an agent can lift.

Why is the v0.2 trust model the interesting part?

Version 0.2, shipped weeks after 0.1, is where OKF stops being a filing convention and becomes a position on the biggest open question of the agentic web: when most content is machine-written and machine-read, how does anyone know what to trust? The spec's answer is to make trust inspectable rather than asserted. Every concept can carry sources, the materials it derives from, each with objective credibility signals (who authored it, when it last changed, how heavily it is used), and individual claims in the body attribute themselves to those sources through keyed footnotes. Authorship is split from confirmation: generated records who or what wrote the content, while verified records who checked it, and the actor convention distinguishes agents, automated processes and people. From that, consumers derive three trust tiers: unverified, machine-confirmed, and human-reviewed, with the human prefix as the top tier. Freshness is a plain comparison against an absolute stale_after instant, and lifecycle status marks drafts and deprecations.

The most forward-looking piece is the Attested Computation: a concept that carries not just what a number means but the sanctioned way to compute it, with a deterministic, no-LLM attester that can confirm a given run really executed the blessed computation rather than something an agent improvised. Verification confirms the definition still matches policy; attestation confirms each individual number was produced the approved way. That is a governance vocabulary the open web does not have yet, and it is telling that the first serious attempt at one arrived from the agent infrastructure side rather than the search side.

What does OKF mean for modern search?

Here the honest answer has two halves, and both matter. The first half: OKF is not a search standard. It comes from Google Cloud, not Google Search; the specification contains no discovery mechanism, no well-known URL, and no crawler commitment; and as of today no search or answer engine fetches OKF bundles from websites. A bundle will not move rankings or AI visibility this week, and any agency selling OKF as a visibility tactic is ahead of the evidence. In Laurelin's audit register it sits in the same evidential position as llms.txt, unfetched, with one decisive difference: llms.txt is an unofficial proposal with no steward, while OKF is a versioned specification with reference implementations, a consuming ecosystem in agent tooling, and an owner with obvious reach into the crawler side of the same company.

The second half is the structural signal. Search is visibly decomposing into retrieval for machines that answer on your behalf: AI Overviews and chat assistants today, task-executing agents next. Those systems need exactly what OKF encodes and what HTML does not: which claims derive from which sources, who verified what and when, whether a fact has gone stale, and which computations are sanctioned. Today, answer engines approximate all of this from on-page signals, the byline, the linked statistic, the schema graph. A format that makes the same properties machine-checkable is the plausible next layer, and if a discovery convention ever lands, a well-known path or a link relation an ecosystem actually fetches, sites with curated bundles will be a one-line change from participating while everyone else starts from zero. That asymmetry, tiny cost now against a plausible option later, is the whole strategic case.

What does it mean for content teams and publishers?

The practical reframe is that OKF treats content as knowledge rather than as pages. A page is a persuasive artifact: narrative, brand, design. A concept file is the same substance stripped to its load-bearing facts, typed, sourced, dated and linked to its neighbours. Teams that already run structured editorial operations, defined topics, named authors, sourced claims, review workflows, will find that OKF is essentially their editorial governance made machine-readable: the review step becomes a verified entry, the style guide's sourcing rule becomes the sources family, the content refresh calendar becomes stale_after. Teams whose content would embarrass them in that format, unsourced claims, no named authors, no review trail, have discovered something about their content, not about the format.

There is also an immediate private payoff that requires no one else to adopt anything: a curated bundle of your own expertise is directly useful to your own AI tooling, internal assistants, drafting agents, client-facing chat, because it solves your own context assembly problem. That is the correct near-term business case, with public discovery as the option it keeps open. The one discipline that keeps the public version safe is mirroring: the bundle must never say more than, or differently from, the visible site. A knowledge layer that diverges from the rendered pages is the same class of risk as cloaking, detectable by the same diff, and the trust metadata makes any dishonesty attributable.

How does OKF compare with llms.txt and Schema.org?

The three are not competitors; they stack. Schema.org tells engines what a page is. llms.txt tried to tell language models where to look and never earned a consumer. OKF specifies what curated knowledge looks like once something does come looking, and the source it derives from, on a well-run site, is the page that Schema.org already describes.

What should you do now?

Three moves, in order of certainty. First, keep winning the layer that is actually fetched: the visible page and its structured data, built to the standard set out in the perfect HTML page guide, because every current answer engine judges you there. Second, publish a mirroring bundle at /okf/, generated from your pages, with honest generated and verified metadata and stale_after on anything dated, and expose it with a link rel="alternate" type="text/markdown" in the head plus a visible footer link, since a plain anchor is the one discovery mechanism everything already understands. Third, put the same bundle to work internally, where it pays back immediately. And keep the claim honest with clients and stakeholders: this is an option on the agentic web, purchased for almost nothing, not a ranking factor, and the moment that changes the evidence will say so before the invoices should.