What Is OKF (Open Knowledge Format)? A Plain-English Guide for 2026

What is the Open Knowledge Format?

The Open Knowledge Format (OKF) is an open, vendor-neutral specification for writing down what your organisation knows, in plain Markdown files that an AI agent can read directly. Google Cloud published version 0.1 on 12 June 2026, authored by Sam McVeety and Amir Hormati from the Data Cloud team.

An OKF bundle is a directory of Markdown files. Each file describes one concept, and opens with a small block of YAML frontmatter. Version 0.1 requires exactly one field of every concept: type. There are a handful of reserved fields you can use for querying — title, description, resource, tags and timestamp — and two reserved filenames: index.md, which acts as the entry point so an agent can work its way down the hierarchy, and log.md, which keeps a chronological record of changes.

That is the format. No compression scheme, no new runtime, no SDK to install. If that sounds underwhelming, good — that is deliberate. Somewhere right now a conference speaker is finishing a slide deck titled “OKF: The Death of SEO”, and they are wrong.

The problem it was built to solve

Most of what an AI model needs in a business setting is not on the public internet. It is internal: the meaning of a metric, the runbook for when the payment provider goes down, the reason you stopped offering a service in 2024. That knowledge is scattered across wikis, spreadsheets, Slack threads and one person’s head.

The result, in Google’s words, is that “every agent builder is solving the same context-assembly problem from scratch.” Every new tool re-reads the same documents, badly, and reaches its own slightly different conclusion.

OKF flips it: you state the knowledge once, in a machine-first format, and every agent reads the same canonical source.

The bit the industry is already getting wrong

Here is the part you will not see in the breathless LinkedIn posts: OKF is not a search ranking signal.

Google positioned it as infrastructure for AI agents and internal knowledge, not as a way to climb the results page. There is no evidence it influences organic rankings, and Google has not suggested it does. Search Engine Journal, covering the launch, noted the same thing.

Google were also unusually candid about its maturity, stating plainly: “OKF v0.1 is a starting point, not a finished standard.”

We flag this for the same reason we tell clients the truth about llms.txt: the fastest way to waste a budget in 2026 is to buy an emerging file format from someone selling it as a ranking hack. If an agency offers to “implement OKF to get you into ChatGPT”, ask which line of the specification says that. There isn’t one.

So we could stop here and tell you to ignore it. That would be the wrong call.

So why should you care?

Because the discipline OKF encodes is the same discipline that decides whether AI search can understand your business at all.

Generative engine optimisation — getting cited by ChatGPT, Google AI Overviews, Gemini and Perplexity — rewards clarity and consistency far more heavily than traditional SEO ever did. Answer engines are not ranking blue links for a human to sift through. They are picking one answer, and to be picked, a machine has to understand what you do, where, for whom, and be confident enough to put its name to recommending you.

Everything OKF asks for pushes that way: one canonical version of a fact instead of six contradictory ones, explicit typing, cross-links so knowledge forms a graph rather than a pile, and timestamps so staleness is visible. Do that work and it pays off where it is proven to matter — your schema markup, service pages, FAQs and llms.txt. The bundle is a by-product; the thinking is the asset.

There is a nearer-term win too. If you run any AI assistant — a website chatbot, an internal ops agent, a tool that drafts your quotes — it is only as good as the context you hand it. Most are improvising from whatever they scraped off your homepage. A curated bundle fixes that today, with no reliance on anyone’s roadmap.

How to plan a bundle for your business

Resist the urge to open a code editor. The hard part is deciding what is true.

Start with the questions you answer twenty times a week. What is your lead time? What areas do you cover? What does the deposit actually include? What happens if a customer cancels? If a new starter would need to ask, it belongs in the bundle.

Pick one owner per concept. The failure mode of every knowledge base is drift: two versions of the same fact, both stale, neither wrong enough to notice. One named human per concept fixes most of it.

Type things honestly. A price is not a policy. A service is not a case study. The typing is what lets an agent route a question correctly instead of confidently reciting your refund policy at someone asking about van hire.

Decide what is public and what is not. This matters. An OKF bundle is designed for internal knowledge, and internal knowledge includes things you would not want quoted back to you by a competitor’s chatbot. Margins, supplier terms and staff notes stay behind the wall.

How to actually build one

Once the thinking is done, the build is genuinely small.

1. Create the directory. One folder, an index.md at the top, and a Markdown file per concept. Nest by theme, with an index.md at each level.
2. Write the frontmatter. type is the only field version 0.1 demands. Add title, description, resource, tags and timestamp where they earn their place.
3. Cross-link generously. Each concept should point to three to six genuinely related ones. If a concept links to nothing, it is probably orphaned in your real business too.
4. Generate it, don’t hand-write it. This is the one we would push hardest. Hand-written knowledge files drift from the website within a month. Generate the bundle from the same content model that builds your pages, so it rebuilds on every deploy and cannot go stale. We build client bundles this way for exactly that reason.
5. Keep it out of the search index. It mirrors your page content, so noindex it and leave it out of your sitemap. Duplicating your own site against yourself helps nobody.

Common mistakes to avoid

• Buying it as an SEO product. It isn’t one. Anyone telling you otherwise is guessing or selling.
• Hand-maintaining the files. Drift is guaranteed. Automate or don’t bother.
• Publishing your internals. Read every file as though a competitor will.
• Doing OKF before the basics. If your site is slow, thin, or has no schema, fix that first. A tidy knowledge bundle bolted to a weak website is a beautifully organised filing cabinet in an empty office.

Frequently asked questions

Will OKF help me rank on Google?

No. Google published it for AI agents and internal knowledge, and has not presented it as a ranking factor. Treat any claim otherwise with suspicion.

Is it the same as llms.txt?

No, and they point in opposite directions. llms.txt is a single file on your domain aimed outward at AI crawlers. An OKF bundle is a directory of curated knowledge aimed inward, at agents working on your behalf. They complement each other; neither replaces the other.

Should I implement it now or wait?

If you already run AI tooling, the curation work pays for itself immediately. If you don’t, do the thinking now and ship the files later. Version 0.1 will change, and the notes will outlive the format.

The honest conclusion

OKF is a small, sensible specification that solves a real problem, arriving in a market extremely keen to misrepresent it as a growth hack. It will not lift your rankings. It will not talk ChatGPT into recommending you.

What it will do is force you to write down what your business knows, in one place, in a form both a person and a machine can read. Businesses that do that tend to do better in AI search — not because of the file format, but because they are finally legible. The format is just the nudge.