What Google's Open Knowledge Format (OKF) Means for Marketers

The Open Knowledge Format is an open spec, published by Google Cloud on 12 June 2026, for representing curated knowledge as a directory of markdown files with YAML frontmatter. Every concept is one file. The only required field is its type.

Sam McVeety and Amir Hormati, two engineering leads on Google's Data Cloud team, designed it to fix a data-warehouse problem. Tables and datasets carry no context, so the agents querying them guess. OKF gives each concept a plain-language file an agent can read, published as an open standard.

The format is deliberately small. A file holds YAML frontmatter and a markdown body, and concepts link to each other with ordinary markdown links, which turns a folder into a graph. Per the v0.1 spec, only type is required. Five more fields are recommended, not mandatory, and producers can add any others they want.

One of Google's three design principles is "format, not platform." OKF is not tied to any cloud, database, model provider, or agent framework. In Google's words, it "will never require a proprietary account or SDK to read, write, or serve." The team shipped reference tools to prove it: an agent that documents a BigQuery dataset, a single-file HTML graph viewer, and three sample bundles built from GA4 e-commerce, Stack Overflow, and Bitcoin public datasets.

A marketing team owns the same kind of asset a data team does: curated knowledge that other systems need to read correctly. Your brand voice, your ICP, your positioning, your approved claims, and your product facts are all concepts in the OKF sense. Most of them live in slide decks, wikis, and the heads of three senior people.

OKF formalizes a pattern the researcher Andrej Karpathy named the LLM wiki. Karpathy, quoted in Google's announcement, made the case in one line: "LLMs don't get bored, don't forget to update a cross-reference, and can touch 15 files in one pass." A model can read a hundred small files faster than a person can open one. For years, that knowledge was never written down in a form a model could read.

Reframe your brand guidelines as a directory of typed markdown files and the parallel is exact. One file per claim, one file per persona, one file per product. Each carries a type, a one-line description, and a link to its source of truth. This is the same move I described in agent-addressable content and context engineering: brand knowledge an agent can address directly, not prose buried in a PDF.

Start with the internal case, because it pays off first. If you run AI agents for research, drafting, or social, each one needs your brand context to produce on-brand work. When that context is scattered across tools, every agent reads a slightly different version, and the output drifts.

This is the discipline Anthropic calls context engineering: "curating and maintaining the optimal set of tokens" an agent sees at inference time. An agent acts on whatever context it is given.

A single OKF-style directory fixes the drift at the root. One folder holds the canonical voice file, the ICP files, and the claim files, and every agent reads the same source. When positioning changes, you edit one file, and the next run of every agent reflects it. This is what my AI marketing framework means by keeping the context portable. The file is the system of record, and the agents are interchangeable.

The external case is bigger, and you control less of it. AI assistants now answer questions about your brand directly, and at real scale. ChatGPT alone reached an estimated 800 to 900 million weekly users in 2025, by Andreessen Horowitz's count. When someone asks an assistant what your product does, the model answers from whatever it can read.

Often it reads wrong. In Skyword's June 2026 buyer research, when AI-generated information conflicted with a brand's own messaging, only 12% of people trusted the AI answer, and 54% went looking for outside sources to compare. Nearly one in five had already avoided a purchase based on what an assistant told them about a brand.

Publishing agent-readable brand knowledge is how you shape that answer. The academic version is Generative Engine Optimization, defined in a 2024 paper that showed structured, well-sourced content can lift a source's visibility in AI answers by up to 40%. The web already has lighter versions of the idea: schema.org structured data, and the proposed llms.txt file that hands LLMs a clean markdown summary of a site. Of the three, OKF goes furthest: it represents a whole graph of linked concepts, not just one page's markup or a site summary. Measuring whether it works is a share-of-model problem.

The strategic reason to adopt an open format is ownership. When your brand's knowledge lives inside one vendor's platform, you do not really own it. You use it on their terms, and only while you pay. If the contract ends or the tool drops the model you depend on, the knowledge is stranded. Files you control move with you.

Enterprises already behave this way about models. By Andreessen Horowitz's January 2026 CIO survey, 81% now run three or more model families in testing or production, up from 68% a year earlier. They will not depend on a single provider, and your brand knowledge needs the same hedge.

The value of a knowledge format comes from how many parties speak it, not from who owns it. Google Cloud, on releasing OKF as an open standard.

That is why the consumption side is standardizing too. The Model Context Protocol, an open standard for connecting agents to the systems where data lives, gives any compliant agent a common way to reach your files. Open format on the supply side, open protocol on the demand side. The brands that win the AI-native shift will be the ones whose knowledge is portable enough to feed any model, as I argued in model-agnostic AI marketing. A pile of tools you do not control is still a pile of parts.

You do not need Google's tools or a data warehouse to use any of this. The format is plain markdown and YAML, and the full v0.1 spec fits on a page. Start by writing your brand's Context layer as a small directory of files, one concept per file, and point your agents at it.

1. Pick the five concepts agents get wrong most often: your voice, your ICP, your top three claims, your category definition, and your pricing logic.

2. Write one markdown file per concept. Give each a type and a one-line description, and keep the body short and sourced.

3. Link concepts to each other with plain markdown links, so the folder becomes a graph an agent can walk.

4. Point your marketing agents at the folder as their first context, ahead of any prompt.

5. Publish a public subset for external engines. A clean llms.txt file or schema markup is the entry-level version of the same idea OKF formalizes.

None of this requires permission from a platform. The files are plain text you own, readable by whatever tool you point at them next.