Google’s Open Knowledge Format Could Work For Websites, Too

Google printed a format for turning a physique of data right into a folder of linked markdown information. It was constructed for inside firm information, and by chance, it solves an issue public web sites have too. Proper now, essentially the most an AI agent will get out of your web site is a flat read of your pages, one by one. This format builds a graph of how your concepts join as an alternative, so I attempted it alone web site.

Google’s Open Information Format Is A Listing Of Linked Markdown Recordsdata

On June 13, 2026, Google’s information workforce printed the Open Knowledge Format, or OKF, a option to characterize a physique of data as a listing of markdown information with a skinny layer of YAML frontmatter. Every idea, a desk, a metric, a runbook, an API, will get its personal markdown doc. A brief block of YAML carries the queryable fields, kind, title, description, useful resource, tags, and timestamp; the markdown physique carries the reason, and ideas hyperlink to one another with bizarre markdown hyperlinks, which Google says turns the listing into “a graph of relationships.” There is no such thing as a runtime, no SDK, no construct step. Google describes a bundle in three phrases: “simply markdown,” “simply information,” “simply YAML frontmatter.”

The goal is inside firm information, the context Google says is “locked behind whichever floor created it,” and it’s early, v0.1, which Google calls “a place to begin, not a completed normal.” Nothing within the announcement mentions public web sites. That hole is what this piece is about.

On A Web site, A Information Graph Beats A Flat Web page-Copy

The agent-readable model of your web site, the one a mannequin or a browser really consumes, is flat. Serving every web page as markdown, the best way Cloudflare does on the community edge, is near AMP for LLMs: a second, stripped copy of each web page for a machine to learn. It mirrors what you have already got, web page for web page, and it drops the identical factor each page-by-page copy drops, which is how the pages relate to one another.

A knowledge graph keeps that relationship layer. When your ideas hyperlink to one another, an agent not solely learns what each is, however it additionally learns how they sit relative to one another, which is most of what understanding an internet site really means. Two pages can each point out an idea and by no means inform a machine that one is the framework beneath it and the opposite is the narrower objective beside it. A graph says it outright, in hyperlinks the machine follows. OKF is an off-the-shelf option to construct that graph: markdown, so it’s low cost, and structured, so it carries the relations.

I Tried OKF On The No Hacks Web site

I wrote an OKF bundle for the No Hacks web site, one markdown file every for the model, the host, Machine-First Architecture, the agentic web, Agent Expertise Optimization, Reply Engine Optimization, llms.txt, and WebMCP. Every follows Google’s conventions, the YAML fields on high and a plain markdown physique beneath. The work was largely deciding which ideas mattered and the way they join, not writing the information.

One file, the idea for Machine-First Structure, seems to be like this:

---
kind: framework
title: Machine-First Structure
description: A framework for constructing web sites whose full that means is on the market to a machine studying them, with the human expertise layered on high quite than the opposite means round.
useful resource: https://machinefirstarchitecture.com
tags: [Framework, Machine-First Architecture, Agentic Web]
timestamp: 2026-06-13
--- Machine-First Structure is [Sani](./sani.md)'s framework for the [agentic web](./agentic-web.md). The core thought: construct the content material so a machine studying it will get the entire that means, the info, the construction, the relationships, and the human studying will get that very same that means with the design on high. This is the reason codecs that strip an internet site to plain textual content, like markdown for brokers and [llms.txt](./llms-txt.md), matter. Its functionality facet is [WebMCP](./webmcp.md), and its measurement facet is [Agent Experience Optimization](./agent-experience-optimization.md).

These bracketed hyperlinks on the backside are the graph. An agent following them learns that WebMCP sits below Machine-First Structure, and llms.txt is similar sort of wager, which a flat copy of my pages by no means says out loud. Throughout the eight information, that’s the complete construction: ideas, and the relationships between them.

A bundle like it is a second copy of what the web site already says, and a second copy is a second factor to maintain in sync. The second the web site adjustments, the bundle is mistaken till you replace it too. That tax isn’t distinctive to OKF: it’s what each parallel machine-readable layer prices, an llms.txt file, a markdown mirror of your pages, a bundle like this one. The model an agent reads is simply as correct as your self-discipline in conserving it present.

Google didn’t construct OKF for this. Its goal is inside firm information, and nothing in its plan factors at public web sites, so internet hosting a bundle for a visiting agent is off-label, and it might keep that means. The reader I made it for, an agent that fetches the bundle and follows the graph, would possibly by no means present up. The rationale to do it has to face with out that payoff, and it does: writing the bundle compelled me to state plainly what No Hacks is aware of and the way its concepts join, and that surfaced gaps I’d not have discovered writing one other web page. It’s the identical self-discipline as Machine-First Structure, put your that means in a type a machine can learn, and you discover the place you have been imprecise.

The place A Web site Information Graph Might Lead

None of what follows is a prediction. It’s a course, and it is determined by brokers really studying web site information graphs, which at this time, none do. The form remains to be value seeing.

The identification file might develop right into a information graph. Immediately, llms.txt is a single line saying who you’re. A broadcast bundle is the total model of that concept, a map of every part your web site is aware of and the way the elements join, so the skinny identification layer and the structured information layer turn into one factor.

Brokers might question that map as an alternative of scraping your pages. An agent that pulls your bundle and follows its hyperlinks will get a cleaner, relationship-aware learn than one parsing your HTML one web page at a time, and also you get extra say in how your personal ideas are represented when an AI describes you.

The map might even turn into the canonical layer. The model a machine reads stops being a duplicate of your web site and turns into the supply, with the human pages as one rendering of it. That’s the fully machine-first website the agentic internet has been pointing at, reached via a facet door Google opened for inside information.

Markdown Is Not New

John Gruber created Markdown in 2004, with Aaron Swartz as his beta-tester, and the entire design objective was readability: textual content you’ll be able to learn as-is, with out rendering, that also converts cleanly to HTML. 20 years later, it runs GitHub, Reddit, a lot of the documentation you learn, and the chat containers of the AI instruments themselves. It gained by being legible with out being rendered, which is the precise property that makes it straightforward for a machine to learn.

I’ve written most of what I write in it for 15 years, since iA Author turned my predominant writing app in September 2011, so every week when the agent-readable internet converges on markdown is acquainted floor to me, not a brand new trick. The information behind No Hacks (No Hacks OS mission) has run the identical means for months: markdown information with structured frontmatter, linked to one another, the form a machine can learn and traverse.

Machine-facing formats maintain touchdown on that very same floor, llms.txt, Cloudflare’s markdown, and now OKF. Google itself isn’t of 1 thoughts about it. Its Search facet referred to as llms.txt “purely speculative” for rating, its Chrome side added an llms.txt check to Lighthouse’s agent-readiness audit, and its information workforce has now printed OKF.

If you wish to see the place your web site stands, it takes thirty seconds. Open your most necessary web page and paste it right into a plain-text editor, the place the hyperlinks collapse into plain phrases. Have a look at what’s left and discover something that states how its concepts relate to the remainder of your web site, not that one web page hyperlinks to a different, however the relationship itself. There’s normally nothing, and that absence is what a information graph fills, whether or not or not you ever contact OKF.

OKF is that this week’s information, and the substrate below it, plain textual content a machine can learn, has been right here since 2004. What Google added was a typical and a reputation.

Extra Assets:

This submit was initially printed on No Hacks.

Featured Picture: Roman Samborskyi/Shutterstock