Big Idea

The argument in one line.

Google's Open Knowledge Format replaces the need for custom agent integrations by giving any business a standardized, markdown-based way to make its knowledge directly accessible and exchangeable across AI systems.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

An SEO professional or content strategist who wants to understand how agent discoverability differs from search-engine indexing.
A consultant or expert who sees selling knowledge bundles directly to AI systems as a new revenue stream.
A developer building agentic workflows who wants a structured, compound alternative to naive RAG over raw documents.
An early adopter who prefers a spec walkthrough over wading through Google's GitHub documentation alone.

SKIP IF…

You want step-by-step code — this is a conceptual explainer with no build-along component.
Your work has no overlap with AI agents, knowledge management, or technical content strategy.

TL;DR

The full version, fast.

Google's OKF formalizes the LLM Wiki pattern as a distributable standard: your knowledge becomes a directory of YAML-fronted markdown files, each representing one atomic concept, linked into a traversable knowledge graph. Unlike RAG, which retrieves from raw documents at query time, the LLM Wiki model has the language model incrementally build and maintain the wiki itself, updating entity pages and compounding the synthesis over time. OKF standardizes this so bundles can be shared across organizations, purchased from experts, and consumed by any compliant agent without custom integration.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 01:03

01 · Introduction

Google Cloud Blog announcement framing; why OKF is a new layer for the internet

01:03 – 01:43

02 · How OKF changes SEO

Not about being found — about making businesses accessible to agents; new acronym needed

01:43 – 02:21

03 · Selling knowledge bundles

Proprietary OKF bundles as purchasable assets; experts selling knowledge directly to agents

02:21 – 03:02

04 · How to learn the spec now

NotebookLM + Gemini as fast onramp; prompt: give me 20 ideas of how I could use OKF in my business

03:02 – 04:01

05 · What is OKF?

Markdown files, the LLM Wiki pattern, not taking whole pages but extracting concepts

04:01 – 05:03

06 · Breaking down the spec

Human-and-agent-friendly, minimal format, no schema registry, no central authority

05:03 – 05:51

07 · Terminology

Knowledge Bundle, Concept, Concept ID, Frontmatter, Body, Link, Citation defined

05:51 – 06:30

08 · Concepts as atomic files

One website may produce 10-50 concept markdown files; contrast with page-level thinking

06:30 – 07:38

09 · YAML frontmatter

Required fields: type, title, description, tags, timestamp; agent writes these for you

07:38 – 08:40

10 · Body, citations, cross-linking

Body follows frontmatter; cross-links create the knowledge graph; citations support claims

08:40 – 09:11

11 · Bundle structure and storage

index.md, log.md, subdirectories; store in Git or Obsidian

09:11 – 11:12

12 · Practical examples

BigQuery Customer Orders resource concept; Incident Response Playbook with trigger condition

11:12 – 13:48

13 · Karpathy's LLM Wiki

The conceptual parent of OKF: LLM incrementally builds a persistent wiki vs. RAG at query time

13:48 – 14:34

14 · LLM as author and maintainer

You find and learn; the model writes and maintains; your job is sourcing, not structuring

14:34 – 17:15

15 · OKF as the new schema

Two revenue streams: building OKF as a service; selling knowledge bundles; tool ecosystem emerging

17:15 – 18:05

16 · Agent discovery and semantic unbaking

llms.txt will point agents to your OKF; semantic unbaking as the big shift

Atomic Insights

Lines worth screenshotting.

OKF is not about being found in search — it is about making your business legible to AI agents that act on your behalf.
Each OKF concept is one markdown file representing one idea — not one webpage, not one document, but one atomic unit of knowledge.
The LLM Wiki pattern beats RAG because the synthesis is already done: the model integrates new sources into a persistent wiki rather than rediscovering knowledge on every query.
You never write the wiki yourself — your job is to find and learn; the language model writes and maintains the structure.
OKF bundles may become purchasable assets: buy an expert's curated knowledge graph and integrate it directly into your own agents.
YAML frontmatter (type, title, description, tags, timestamp) is all the metadata OKF requires — your agent writes it for you from the spec.
The log.md file inside a bundle lets agents record what changed and when, making the wiki auditable over time.
Cross-links between concept files turn a flat directory into a knowledge graph — the connections are where the real intelligence lives.
Semantic unbaking is the shift from machine-structured schema back to natural human knowledge that agents can process directly.
OKF creates two new revenue streams: building OKF bundles as a professional service, and selling proprietary expert knowledge bundles to other agents.
Storing your OKF bundle in a Git repository is free, version-controlled, and immediately accessible to any compliant agent.
The format is intentionally minimal: no schema registry, no central authority, no required tooling — if you can read a file, you can read OKF.
A triggered Playbook concept tells your agent exactly where to go in your knowledge when a specific condition fires.

Takeaway

OKF makes your knowledge an asset agents can act on.

WHAT TO LEARN

Google's Open Knowledge Format shifts knowledge from something search engines index to something agents can traverse, reason over, and exchange — and the format is simple enough to start building today.

OKF is not about search ranking — it restructures your knowledge so agents can access, traverse, and act on it without custom integration or API endpoints.
The fundamental unit is the concept, not the page: one markdown file per atomic idea, which forces a clarity about what you actually know that page-level publishing does not.
YAML frontmatter (type, title, description, tags, timestamp) is all the metadata required — the language model writes this for you from the spec, so the barrier to entry is lower than it appears.
Cross-linking concept files is what transforms a flat directory into a knowledge graph — the connections are where agent reasoning compounds.
The LLM Wiki pattern beats RAG for long-lived knowledge because synthesis is already done: the model integrates new sources into the existing wiki rather than rediscovering relationships on every query.
Storing an OKF bundle in a public Git repository makes it immediately accessible to any compliant agent at zero cost — distribution is solved before tooling matures.
The Playbook concept type is particularly powerful: it gives an agent a triggered procedure, replacing ad-hoc prompting with persistent, structured process knowledge.
Owning well-structured OKF bundles is likely to become both a professional service and a direct revenue stream — expertise that was previously locked in documents becomes a purchasable, integrable asset.

Glossary

Terms worth knowing.

OKF (Open Knowledge Format): Google's proposed open standard for representing knowledge as a directory of YAML-fronted markdown files, designed to be authored by people, generated by agents, and exchanged across organizations.
Knowledge Bundle: The distributable unit in OKF: a self-contained, hierarchical collection of markdown concept files. The unit of distribution between organizations or agents.
Concept: A single atomic unit of knowledge within a bundle, represented as one markdown document. May describe a tangible asset, an abstract idea, a process, or anything in between.
Concept ID: The path of the concept file within the bundle with the .md suffix removed. For example, tables/users.md has concept ID tables/users.
Frontmatter: A YAML metadata block delimited by --- at the top of each OKF markdown file, containing type, title, description, tags, and timestamp.
LLM Wiki Pattern: Andrej Karpathy's pattern where the language model incrementally builds and maintains a persistent, interlinked wiki rather than retrieving from raw documents at query time (RAG).
Semantic Unbaking: A phrase describing the shift from forcing human knowledge into rigid machine-readable schema back to natural knowledge structures that agents can interpret directly.
Playbook (OKF type): An OKF concept type that defines a triggered process: when a specific condition is met, the agent follows the steps in this concept file.

Resources

Things they pointed at.

02:36linkGoogle Cloud Blog — Introducing the Open Knowledge Format ↗

03:50linkOKF SPEC.md on GitHub ↗

11:12linkAndrej Karpathy's LLM-Wiki ↗

02:21toolNotebookLM

02:21toolGemini

08:29toolObsidian

18:03linkMarie Haynes Community (The Search Bar) ↗

Quotables

Lines you could clip.

01:08

“This isn't about getting found. Rather, it is a way to make your business accessible to agents.”

Reframes the entire purpose of OKF in one sentence→ TikTok hook↗ Tweet quote

14:07

“You never or rarely write the wiki yourself. The language model is gonna write and maintain all of it. So once you get it going, your job is to find new information.”

Counterintuitive role reversal — agent as author, human as curator→ IG reel cold open↗ Tweet quote

14:34

“I think this is the new schema.”

Short, bold claim that positions OKF in existing vocabulary→ newsletter pull-quote↗ Tweet quote

17:34

“Semantic unbaking — we're actually just gonna be able to have these lives where we learn stuff and we share stuff, and those who work hard to develop expertise will be able to be rewarded.”

Poetic closer with a human payoff→ IG reel cold open↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphor

00:00Google just announced something called the open knowledge format, and boy am I excited about this. This is a whole new layer for the Internet. It's a way to take your knowledge, your business knowledge, your personal knowledge, whatever it is, and structure it so that agents can understand it.

00:16Now this is nothing new. The way that they do it is they structure it in markdown files. It's actually very, very simple.

00:22We'll look at some examples in just a minute. But what is new is that it is a standardized way of doing this so that if my agent wants to access your public OKF, then it can happen standardized.

00:38I don't need to use any particular software to do it. I could use something like Notion or Obsidian to look at the markdown files. I could ask my agent to make a webpage out of it.

00:48I could do all sorts of things, but the key thing is that I have access to the knowledge that a business has chosen to provide. This is a really big deal. I think it's going to change how we do SEO.

01:01I think we're gonna need a different acronym. It's not SEO. It's not GEO.

01:05It's not even agentic search optimization. This isn't about getting found. Rather, it is a way to make your business accessible to agents.

01:16Not just accessible, but make it so that agents can do things with your business knowledge. I do think that people who understand OKF will be in high demand. I think it's gonna be a new service that SEOs offer where kind of like making a site map, but much more detailed, where people will be really good at understanding how to take a business' knowledge and turn it into a knowledge graph that is an OKF.

01:41But I think it's even more exciting for people who actually have proprietary knowledge, processes, or something that people already pay you for because you will be able to I'm fairly certain.

01:52I don't know this a 100% for sure, but you should be able to sell your OKF bundles. And that way, you can have access to an expert's knowledge and even integrate it into your own OKF so that you will have you know, we'll probably buy OKF bundles from a lawyer, from an accountant, from an SEO.

02:15And I I think this is a very, very exciting thing. So this is a proposed standard. It is just the very first version.

02:22And so if you're watching this in the future, I'm recording this mid June, there probably are some new versions that are out there. I'd encourage you to do what I did is take, I've put some links in the description for Google's documentation on this and also the spec file in GitHub.

02:38Put all of that into a notebook l m. Give that notebook to Gemini and then ask Gemini to summarize the information and tell you how you can use this in your business.

02:49One good prompt would say would be to say, give me 20 ideas of how I could use OKF in my business.

02:56So let's get into Google's description of what OKF is. OKF is a specification that formalizes something called the LLM Wiki pattern.

03:05This pattern was described by Andre Karpathy. We're gonna look at that in just a minute. But first, let's look at what an OKF is.

03:13It is just markdown. If you're not familiar with markdown, it's just a really simple way to describe text, you know, as text.

03:22Basically, a text file, it might have a little bit of markup to say this is a heading, this is bolded, or whatever. Just a very simple way to understand text. It is really easy for agents to understand text.

03:33Now, we're not gonna be taking every page of our website and turning that into a markdown file. I mean, we could. We're actually going to take the things that we have on our website, extract the concepts from those, and make those concepts their own markdown files.

03:49Let's see if I can explain that a little more clearly. I found that this was the most helpful document that Google produced. It's on GitHub.

03:56Again, it's linked to in the description, the spec. Md file.

04:01I'm actually going to read a part of this because, uh, it explains things really well. OKF is an open human and agent friendly format, so it's not just for agents. Humans can use it as well for representing knowledge, the metadata, the context, and the curated insight that surrounds data and systems.

04:19It's designed to be authored by people, generated by agents, exchanged across organizations, which I think is the really exciting thing, and consumed by both.

04:29They say that the format is intentionally minimal, and what they mean by this, they're gonna say some words that if you're not into markdown, because I I didn't know what a YAML front matter is, but we're gonna discover that, is actually very, very simple.

04:45There's no schema registry. When I first saw this, my thought was this is just schema. The thing is that it's so simple that you don't need to specifically outline every specific thing that this is an entity, that this is you'll see what I'm talking about.

05:00I think it's important to look at the terminology that Google gives us. We are going to be creating what's called a knowledge bundle. This is a collection of files.

05:10So OKF is essentially a directory of directories, and in those directories are a bunch of markdown files. Nothing new, but again, the fact that it's standardized is the important thing here.

05:21I'm fairly certain that these OKF bundles are things that you could buy or sell. So again, like if I my page on Google algorithm updates and things that changed with AI that might be relevant to rankings, I might sell that as a knowledge bundle that you can purchase and then instead of having to access an MCP for my website or to use WebMCP, you would essentially integrate my knowledge into your system.

05:51Each of the markdown files in these directories represents a concept.

05:57And this is the concept, no pun intended, that's hard to understand here. Because we're used to, as SEOs, taking a web page and saying, alright, here is the summary of the web page.

06:09Instead, each markdown file is a concept, a single unit of knowledge within a bundle represented as one markdown document. So one website might produce maybe 50 or maybe 10, or maybe even more concepts that each would be their own markdown file.

06:27Let's talk about what this YAML front matter is. This is the information that is required in these markdown files.

06:35And then after this information, you can put whatever you want in the file. So you're gonna have the type, the title, the description of here's what's in this file, resource tags.

06:48The tags can be different topics that are covered, and then a timestamp. All of this will be really easy.

06:55Your agent will do it for you. It's not something that you need to really pay a lot of attention to. What you're probably gonna do is give this spec dot m d file to your language model and say, create the YAML front matter for me.

07:08Then, after that are the instructions or the knowledge or the data tables or whatever it is that you want to share on that particular topic. After the front matter, then there is the body that has everything else in it.

07:23And there can also be links. And this is where it's really interesting because eventually what you're creating is a knowledge graph of all of your knowledge, which is really, really interesting.

07:33You can also have citations that help support a claim in the body. You don't necessarily have to have that.

07:39Here's an example of how the bundle can be structured or how it's supposed to be structured. The index file is where every one of these has to have an index file.

07:50This is really interesting, the log file. Log file can be so that your agent says, oh, I updated this today. I did this today.

07:58And we'll look at Carpathi's Wiki in a minute because it's really interesting how the role that your agents play in this. The concept is the most important thing here.

08:08Again, my concept might be WebMCP. And then I could have subdirectories.

08:15Maybe my concept is MCP and I have subdirectories for WebMCP and for other things. This is what's gonna take some time is for us to figure out how to structure these because every business is gonna be different. Then you can store all these files in a git repository.

08:29If you don't know what that is, it's essentially just uploading it to GitHub, totally free to do that. You can also store it in a markdown organizer like Notion or like Obsidian. When this first came out, I thought that it would be a replacement for Obsidian.

08:45I haven't used Obsidian, but here's why I said that. Because I was considering getting something like Obsidian and now I'm like, oh, actually, I can just create my own markdown files.

08:56I could create my own front end to look at those markdown files. I don't think I'm gonna need a way to store and to organize all the markdown files.

09:06But who knows? I could be wrong on that. Now many of the examples that Google gives us are not quite what I'm talking about.

09:13So they talk about using the OKF as a way to get BigQuery data, especially from GA four, which I think is really, really interesting, and it is something that I wanna do. I don't think it's actually where the power is for most of us who are are watching this video.

09:28And this example that they give is it takes a while, at least it took me a while to wrap my head around it. So we have the stuff that needs to happen, that needs to be there at the top, which we just talked about. And then this is the body, which you don't need to have schema in your body.

09:45You don't need to have joins. You don't need to have citations. These are things that this business, this organization is saying, when we look at customer orders, this is how we organize them.

09:56If you didn't have this, if you were just using REG, if you were just using a language model to say, get me the information from BigQuery and organize it, it would guess in how you are going to organize things. This way, you can say, This is how we do it.

10:12This is our way of organizing things. Here is another example and I think this is one that I'm going to use a lot. The type is a playbook, which is kind of like the process.

10:22This playbook would be triggered when a particular thing happens. So mine might be, in my overall OKF of my knowledge, might be a trigger that happens when somebody talks about a traffic drop, And then I would say I have a playbook for diagnosing why your traffic has dropped.

10:41That is the type of thing. So basically, am saying to your agent, here is where you go in my brain to get this information.

10:49And then the interesting thing is the cross linking. And this is something where I think it's going to be very exciting, where we can start to see the connections between things that we've talked about.

10:58We'll know more about this. This will be more interesting when we look at Carpathi's Wiki because I think this is really, really important to pay attention to.

11:07Let's actually look at Andre Karpathy's LLM Wiki. This is not exactly what the OKF is, but I found that reading this helped me understand the power of what we can do with these OKFs a lot better. Karpathy he talks about the idea that most people with when you're using LLMs and you're using documents, you're thinking of rag, that you're giving your LLM this massive amount of context and saying, here's a question.

11:35Go find the parts that are relevant in all of this context in the entirety of your website, and you're relying on your agent to do a lot of work. Whereas, his LLM Wiki is basically saying, here are here's a here's a map to where the important things are in our in our knowledge.

11:54But here's the part that I'm really excited about. In his model, instead of just retrieving from raw documents at query time, the language model incrementally builds and maintains a persistent Wiki. And so he talks about when you add a source, it's not like you're adding a page to a sitemap in your website.

12:14Instead, your language model, I think we're all gonna have to create there probably will be tools that do this for us, but I think the best ones will be ones that we create ourselves. When you add a source, the language model doesn't just index it for later retrieval.

12:29It reads it, extracts the key information, and integrates it into the existing wiki, updating entity pages, revising topic summaries, noting where new data contradicts old claims, strengthening or challenging the evolving synthesis.

12:44So let's say that today, this video goes into my OKF. The topics that I talk about get extracted and each of them get their own markdown file.

12:56Now, let's say that a couple months from now, I do another thing about OKF. Maybe I give an example of how I used it. I did actually create an OKF of our peppers in our garden.

13:07I'll share that at some point. So let's say I share that, then it's not like that's going to create a whole new set of concepts. The concept of OKF has already been discussed.

13:17And so what my language model should do is say, Oh, we already have discussed the idea of OKF. Let s see where I can add to the body in that information. And then it might also discover new links that maybe I have said that OKF is connected to revenue that you can gain from building this type of thing.

13:41Again, more on that in just a minute. In Carpathi's version, says, You never or rarely write the wiki yourself.

13:47This is the part that is hard to wrap my head around, that the language model is gonna write and maintain all of it. So once you get it going, your job is to find new information. Your job is to learn things and to maybe write about them, to share them.

14:03It doesn't even have to be your knowledge. Perhaps when Google publishes new documentation, I put that into my OKF and that becomes, you know, the concepts are Google documentation and, uh, and things like that.

14:16He talks about a few examples here. And I I honestly, I think that it's endless what we could use this for. I think that anytime you wanna take a bunch of knowledge and you wanna do stuff with agents, this is what we're gonna use.

14:27So let's talk about where this is going because I think that this is the new schema. I I think it's funny because I have historically really struggled with understanding I understand the importance of some schema.

14:41If you have an ecommerce site and you have rapidly changing product prices and things, schema helps the the the web, helps Google, helps agents understand rapidly changing things on your website. But I didn't like the idea of mapping out every individual entity on your site, every piece of schema. I just didn't understand that.

15:04And the reason is that the web is not our knowledge is not like that. It's not natural. This, what Google has described here, is a natural way to say, here's what I learned.

15:16Now go categorize it into my brain. And I think that this is gonna evolve over time. Again, this is just the very early first version.

15:24I do think that we will have two revenue streams here for most of you who are watching this video. One is selling this service to businesses. There will be tools.

15:36I know Seganthan Mahanadasan has created already a tool that will take your web pages and turn them into an OKF bundle. However, he is, I think, just taking each web page.

15:48He's not actually doing the concepts. I'm sure that that will come. I'm sure there will be many tools.

15:52But I think that there will be great skill in knowing individually and understanding a business, in understanding how to organize that knowledge that is going to be very, very valuable. So I would encourage you to play with this, to create an OKF of just something.

16:09It doesn't have to be your entire website, but just something so that you can use it. The next thing is that this will be valuable to people who have knowledge to sell. And I do think that we're going to sell OKF bundles to each other.

16:23I think that it it kinda blows my mind at how all of a sudden, I won't just have an MCP where I can get legal information to review a document. Like, I could have my lawyer or the best lawyer in this area, you know, have his or her OKF and and have that as part of my business intelligence.

16:45It is really hard to to wrap my head around that. And I think and then not only that, it's not like I'm just having their files and their static knowledge. But as the law changes, then what I need to know for my business knowledge gets updated as well.

16:59We're gonna build this whole entire network of a brain where those who have knowledge and expertise in an area will be very, very valued. Another question that I saw come up is how will agents know that you have this OKF? And I actually think it's gonna be in your llms.

17:17Txt that there will be a direction to agents to say that you have an OKF bundle. We'll see what happens with that.

17:25I'm gonna end with this phrase that Gemini gave me. I've spent endless hours brainstorming on what OKF means. And Gemini gave me the phrase semantic unbaking, and I think that that's just amazing.

17:36I think that we're going to be able to take our knowledge and instead of having to work like machines and make it into something that is machine readable, we're actually just gonna be able to have these lives where we learn stuff and we share stuff, and those who work hard to develop expertise will be able to be rewarded and will be able to help a lot of people with this.

17:58I hope you found this interesting. I will be writing more about OKF in my community, community.mariehaines.com. And, uh, thanks so much for watching.

18:07I wish you the best of luck with every

The Hook

The bait, then the rug-pull.

Google quietly dropped a spec that could change what it means to publish knowledge on the web. Not for crawlers this time — for agents. The Open Knowledge Format is a directory of markdown files, a handful of YAML fields, and a radical bet that the next layer of the internet is built from atomic concepts, not pages.

Frameworks

Named ideas worth stealing.

05:07concept

Knowledge Bundle

The distributable unit of OKF: a directory tree of concept markdown files that can be stored in Git, shared, or sold.

Steal forPackaging any proprietary process or expertise into an agent-consumable format

11:12model

LLM Wiki Pattern (Karpathy)

Instead of RAG, the LLM incrementally builds a persistent, interlinked wiki. New sources are read, integrated, and contradictions resolved, compounding over time.

Steal forBuilding any long-lived agent knowledge base that should compound rather than retrieve

06:30list