Essay · 04

Google's new knowledge format is real. Three of the loudest claims about it aren't.

· 6 min read

Five days ago, on 12 June 2026, Google Cloud published something called the Open Knowledge Format — OKF for short. A way to write down what your business knows in plain files an AI agent can read and act on (Google Cloud's announcement; the spec on GitHub).

I watched the first wave of explainers go up about it, and I went and did the thing I'd tell any client to do: I read the actual spec instead of the hot takes. The two don't match. The explainers are already selling a future the document doesn't describe.

So this is me showing my working. What OKF genuinely is, what it isn't, and the one question about it that matters most to anyone in my line of work — which almost nobody is asking.

What OKF actually is

Here's the deal, with the jargon stripped out. OKF is a standard way to file your knowledge as a folder of plain text. Markdown files — the same simple text-with-light-formatting these notes are written in. Google calls a folder of them a knowledge bundle. Inside, each file holds one concept: one table, one process, one definition, one playbook. Not a copy of a web page — a single unit of knowledge.

Each file opens with a short block of labels (the spec calls it front matter). Only one label is actually required: type — what kind of thing this is. Title, description, tags, a timestamp — all recommended, none mandatory. After the labels comes the body: the knowledge itself, in whatever shape it needs. Files link to each other with ordinary markdown links, and those links form a map — a graph of how your knowledge connects. There's an optional index file at the front and an optional log file that records what changed and when.

That's it. The spec is about 450 lines, and Google is blunt about its status: it's "Version 0.1 — Draft," and in their words, "a starting point, not a finished standard" (SPEC.md).

One thing worth getting right, because people are already overstating it: this is an open spec published by Google Cloud engineers, released under an open licence. The repository itself carries Google's standard "this is not an official Google product" line. So it's real, and it's from Google Cloud — but it isn't a supported, shipping Google product you can buy. It's a proposal, deliberately put out in the open.

The Karpathy part is real

The explainers tie OKF to Andrej Karpathy's "LLM Wiki" idea, and that connection holds up — Google cites and quotes him directly in the announcement. Karpathy's point, from his own write-up earlier this year, is that instead of an AI re-reading all your raw documents and rediscovering the answer from scratch every single time you ask, it should maintain a living wiki: read new sources once, file what matters into the right pages, update cross-references, note where new facts contradict old ones (Karpathy's original gist).

Worth a caveat the headlines skip: Karpathy frames the wiki as working alongside the old retrieve-on-the-fly method, not replacing it. The "this kills RAG" framing is something added on top by other people, not what he said.

That's the honest version of the idea. It's genuinely good. Now the part where the explainers run ahead of the document.

Three claims the spec doesn't support

I grepped the actual spec for each of these. Here's what's in it.

Claim 1: "Agents will discover your knowledge through your llms.txt file." The spec mentions no discovery method at all. None. The word "llms.txt" does not appear in it. How an agent is supposed to find your knowledge bundle in the first place is simply not addressed. That's not a detail — I'll come back to it, because it's the whole game.

Claim 2: "You'll buy and sell knowledge bundles in a marketplace — purchase a lawyer's expertise, plug it into your own." There is no marketplace in the spec. No buying, no selling, no commercial mechanism of any kind. It's a file format, the same way a PDF is a file format. You could sell files — you can sell anything — but the idea that OKF comes with a built-in market for expertise is invented. It's a nice story. It's not in the document.

Claim 3: "OKF is the new schema — it replaces schema.org, MCP, RAG." The spec doesn't mention schema.org, doesn't mention MCP (the Model Context Protocol, the standard for plugging tools into AI), doesn't mention RAG. It explicitly says it does not replace existing data schemas — it references them. It's not positioned against any of those things. The "new schema that kills the old stack" framing is pure positioning by people who need a bigger headline than "Google published a tidy file format."

None of that makes OKF unimportant. It makes the people explaining it unreliable. Which, given what I do, is rather the point.

The question nobody's asking

Here's the one that actually matters to me, and to anyone whose business depends on being found by AI. OKF has no discovery layer. You can author a beautiful, well-organised knowledge bundle about your business, and as the spec stands today, nothing tells an AI agent it exists or where to look. OKF describes how to write your knowledge for an agent that's already reading it. It says nothing about how the agent shows up at your door.

So read carefully, because this cuts against the breathless version: as it stands, OKF is not a get-found play. It's a get-understood play. It's for the moment after an agent already has access to your knowledge — making sure it reads it cleanly instead of guessing. That's valuable. It is not the same thing as visibility, and anyone selling it to you as "the new way to get cited in AI answers" is skipping the step where the agent finds you at all.

That gap will get filled — by Google's own catalogue and agent tooling, by some convention, by something. But it isn't filled now, and pretending otherwise is exactly the kind of thing that gets people burned.

There's a tell in how Google's playing it, too. They gave the format away for free — because the format was never the scarce part. What Google sells is the layer underneath: the catalogue, the agent gateway, the storage and access and audit. The free spec quietly points demand at the part Google charges for (good analysis here). That's not a criticism. It's just worth seeing clearly before you build a business on the giveaway.

So what should you actually do

If you run a business and you're wondering whether to care: yes, but calmly, and for the right reason.

The format is the easy part — it's a folder of text files, and there will be a dozen tools that spit one out from your website inside a month. The early ones already do, and they do the dumb version: one file per web page. That misses the entire point. The skill — the part that won't get automated away for a while — is judgment. Deciding what counts as a single concept. Knowing which fifteen things your business actually knows that an agent needs cleanly separated and cross-linked, versus the four hundred pages of marketing that should collapse into them. Structuring a body of expertise so a machine reads it the way you'd explain it to a sharp new hire.

That's not a file-conversion job. That's the same work I already do — deciding what to say, in what order, so the thing that gets read and reused is yours. OKF just gives it a tidy container.

So the move right now isn't to rush a knowledge bundle onto your site and wait for the citations. There's no mechanism for that yet. The move is to understand the format, build one for something real to learn how it behaves, and be ready — because the businesses whose knowledge is already cleanly organised will be the ones that drop straight in when the discovery layer arrives.

Who this is and isn't for

If you want a tool that converts your site into the new format and a promise that AI will start quoting you next week, save your money. The tool exists; the promise is fiction. There's no path from "I have a knowledge bundle" to "I get found" in the spec today.

If you're an operator with real, hard-won expertise — the kind people already pay you for — and you want it organised so that when agents do start reading business knowledge directly, yours reads as authoritative instead of as a scraped pile, this is worth your attention. Not your panic. Your attention. That organising — turning what you know into something a machine reads as expertise — is the work I do. I ghostwrite the content that gets cited by AI search. OKF is a new container for it. The judgment that fills the container is the part that was always the job. If you want that done on your category, the contact form is the channel.

What I'm deliberately not claiming

Because receipts matter, and so does saying where the evidence stops:

The format is real. The judgment is still the job. Everything in between is people getting ahead of a five-day-old draft.

← All essays