Graph RAG Meets the Open Knowledge Format (OKF) — A New Foundation for Generative Engine Optimization
For most of the last two decades, search engine optimization revolved around a stable unit of value: the web page. Pages were crawled, indexed, ranked, and clicked. Optimization meant writing better pages, earning links to pages, and structuring pages so crawlers could understand them.
That era is ending. Two developments — one from the retrieval side of AI, one from the publishing side — are quietly redefining what “content” even means in an agent-driven world.
The First Shift: Graph RAG and the Rise of Structured Retrieval
Traditional Retrieval-Augmented Generation (RAG) breaks documents into chunks, embeds them as vectors, and retrieves the closest matches to a user’s query. It works — but it has a well-documented weakness: semantic loss. Relationships between entities, cross-references between concepts, and the broader context that gives information meaning are often stripped away when text is chunked and vectorized.
Graph RAG is the response. Instead of treating knowledge as a bag of chunks, it represents knowledge as a graph of entities and relationships. AI systems can then traverse this graph, reasoning across connected concepts rather than pattern-matching isolated snippets. The result is more coherent, more accurate, and — critically for agents — more actionable retrieval.
The Second Shift: The Open Knowledge Format (OKF)
In June 2026, Google introduced the Open Knowledge Format (OKF) — a vendor-neutral specification for representing knowledge as a directory of Markdown files with YAML frontmatter. As Sam McVeety and Amir Hormati of Google Cloud describe it, OKF formalizes the “LLM-wiki” pattern into a portable, interoperable format that “lets wikis written by different producers be consumed by different agents without translation.”
The design is deliberately simple:
- Just Markdown — readable in any editor, indexable by any search tool
- Just files — shippable as a tarball, hostable in a git repo
- Just YAML frontmatter — for the small set of structured fields agents need to query
But the implications are anything but simple. Marie Haynes, reflecting on Google’s announcement, predicts that OKF “may change how we approach SEO. We will shift from working to be found by search engines to making business knowledge accessible so agents can perform tasks with it.” She goes further, suggesting the emergence of a new discipline — one that isn’t quite SEO, GEO, or even Agentic Search Optimization — centered on turning “a company’s mess of data and processes into a clean knowledge graph.”
Why These Two Shifts Matter Together
At first glance, Graph RAG and OKF address different problems. Graph RAG is about how AI retrieves knowledge. OKF is about how organizations publish it. But they share a common assumption that upends traditional SEO: that the atomic unit of knowledge is not the page, but the concept — a discrete, self-contained node that can be linked, traversed, and reasoned over.
As Haynes notes, when building an OKF bundle, “a single webpage may end up producing quite a few concept markdown files.” Each file represents one unit of knowledge. Google’s spec reinforces this: an OKF bundle is a directory of concept files, each with a small block of structured metadata and a Markdown body describing “anything you want to capture, including tables, datasets, metrics, playbooks, runbooks, and APIs.”
This is the same atomic unit Graph RAG operates on at the retrieval layer. One publishes concepts; the other retrieves them. Together, they suggest a future in which visibility is determined not by how well your pages rank, but by how cleanly your knowledge decomposes into navigable, agent-consumable nodes.
The Thesis: A New Foundation for Generative Engine Optimization
The purpose of this article is to explore that convergence. We’ll examine:
- What Graph RAG actually does, and why it represents a departure from chunk-based retrieval
- How OKF works as a publishing standard, and what makes it different from existing content formats
- The hidden relationship between the two — how they form complementary layers of the same emerging AI knowledge stack
- What this means for Generative Engine Optimization (GEO): what to optimize, what to publish, and what new skills the discipline will demand
As Haynes puts it, “SEOs who can make good OKF bundles will be in demand.” The same is likely true for anyone who understands how Graph RAG systems consume structured knowledge.
Graph RAG: From Chunks to Knowledge Graphs
Graph RAG combines two established technologies into a single retrieval architecture:
- Entity knowledge graphs — structured representations where nodes are entities (people, places, concepts) and edges represent the relationships between them.
- Retrieval-Augmented Generation (RAG) — a process where a large language model is grounded with external knowledge retrieved at query time.
The difference from traditional RAG is fundamental. Instead of fetching loosely related text chunks based on vector similarity, Graph RAG retrieves interconnected concepts — providing the LLM with both the facts and the relationships that bind them together.
The underlying rationale comes from Google’s paper “Relational inductive biases, deep learning, and graph networks,“ which argues that modeling the world as a system of entities and relationships dramatically improves the compositional structure of learned knowledge. Put plainly: AI reasons better when it sees a network of meaning, not a pile of paragraphs.
The Key Components of a Graph RAG Pipeline
Microsoft’s “From Local to Global” paper outlines a canonical Graph RAG pipeline in five stages. Each stage transforms unstructured text one step closer to a reasoning-ready knowledge graph.
Stage 1: Source Documents → Text Chunks Documents are split into manageable chunks. This step balances two competing needs: high recall (catching every entity mention) and efficiency (minimizing the number of LLM calls required downstream).
Stage 2: Text Chunks → Element Instances An LLM reads each chunk and extracts entities, relationships, and additional attributes, producing structured tuples. These tuples become the raw nodes and edges of the knowledge graph.
Stage 3: Element Instances → Element Summaries Each node and edge is summarized into a descriptive text block — an abstract, semantic representation of what that entity or relationship means in the context of the source material.
Stage 4: Element Summaries → Graph Communities The graph is partitioned into hierarchical communities using algorithms like Leiden, which groups nodes based on the modularity of their connections. This creates layered clusters — from broad thematic communities at the top of the hierarchy down to specific subtopics at the bottom.
Stage 5: Community Summaries → Global Answer When a user query arrives, the system pulls summaries from the relevant communities and generates partial answers in parallel. These partial answers are then aggregated into a single, comprehensive response that reflects both local detail and global context.
Why This Matters for Content and Knowledge
Two implications stand out for anyone thinking about how AI systems consume content:
First, structure is created whether or not you provide it. Graph RAG pipelines will extract entities and relationships from your prose regardless of how well (or poorly) that structure is expressed. But the quality of the resulting graph depends heavily on the clarity of the source. Ambiguous entity references, unclear relationships, and buried context all degrade the graph.
Second, the retrieval unit is no longer the chunk — it’s the community. Graph RAG doesn’t answer a query by finding the “best chunk.” It answers by identifying the relevant neighborhood of the knowledge graph and synthesizing across it. This means content that lives as an isolated node — disconnected from related entities and concepts — is far less likely to contribute to an answer than content that’s woven into a dense web of relationships.
The Connection Point to OKF
This is where the bridge to the Open Knowledge Format becomes visible. Stages 1–3 of the Microsoft pipeline exist to reconstruct structure that was lost when knowledge was originally written as prose. OKF short-circuits those stages by asking producers to publish structure directly: one concept per file, explicit metadata, declared relationships.
In a Graph RAG world, OKF is not just a convenient publishing format — it’s a way to skip the lossiest part of the pipeline entirely.
The Open Knowledge Format (OKF): A New Standard for Agent-Ready Knowledge
In June 2026, Google Cloud engineers Sam McVeety and Amir Hormati introduced the Open Knowledge Format (OKF) — a vendor-neutral specification for representing knowledge in a way that both humans and AI agents can consume.
The problem OKF solves is one that had been quietly emerging inside enterprises building agentic systems. Teams were increasingly creating what McVeety and Hormati call “LLM-wikis” — collections of Markdown files stored in git repos, describing datasets, metrics, playbooks, runbooks, and APIs in a way agents could reason over. The pattern worked, but every team invented its own conventions. Wikis written by one team couldn’t be consumed by another team’s agents without translation.
OKF was designed to end that fragmentation. As the authors put it: it “lets wikis written by different producers be consumed by different agents without translation.”
The Design Principles
OKF is deliberately minimal. Its design rests on three commitments:
1. Just Markdown. Every concept is a Markdown file. That means it’s readable in any text editor, renderable on GitHub, and parseable by any tool that understands Markdown. No proprietary formats, no special software required.
2. Just Files. An OKF bundle is a directory of files. It can be zipped, tarballed, committed to a git repo, or mounted on a filesystem. There’s no database dependency, no server requirement, no API to negotiate.
3. Just YAML Frontmatter. Each concept file begins with a small block of YAML frontmatter — a structured header that exposes the fields agents need to filter, query, and connect concepts. The body below the frontmatter is free-form Markdown.
The philosophy is clear: maximize interoperability by minimizing surface area. OKF doesn’t try to model everything. It provides a lean core spec and lets producers extend it as needed.
The Anatomy of an OKF Concept File
A single OKF concept file typically contains:
type— the kind of concept being described (e.g., dataset, metric, table, runbook, API)title— a human-readable namedescription— a short summarytags— freeform labels for classification and filteringresource— pointers to underlying assets (a database table, an API endpoint, a document)- Markdown body — the free-form explanatory content
McVeety and Hormati emphasize that a concept can be “anything you want to capture, including tables, datasets, metrics, playbooks, runbooks, and APIs.” The format is intentionally broad, because the goal is to give agents a consistent way to consume whatever knowledge an organization needs to expose.
Bundles: The Unit of Distribution
Individual concept files are useful, but the real unit of distribution in OKF is the bundle — a directory containing many concept files, often organized into subdirectories by topic or domain.
Marie Haynes, walking through the practical implications, observes that when you take existing web content and convert it to OKF, “a single webpage may end up producing quite a few concept markdown files.” A product page might decompose into separate concepts for the product itself, its specifications, its pricing tiers, its use cases, and its integrations. Each becomes a first-class node — individually addressable, individually citable.
This is a profound shift from the traditional web publishing model, where a page is a single indivisible artifact. In OKF, knowledge is broken down to its atomic components before it’s ever published.
Who Publishes OKF Bundles — and Why
OKF is designed for a range of producers:
- Internal enterprise teams publishing knowledge for their organization’s own agents
- Data providers exposing datasets and metrics to downstream consumers
- SaaS vendors describing their APIs and capabilities so agents can invoke them
- Content publishers and brands — the group Marie Haynes focuses on — making their business knowledge accessible to external agents like ChatGPT, Gemini, and Perplexity
For the last group, the incentive is visibility. As Haynes puts it, OKF “may change how we approach SEO. We will shift from working to be found by search engines to making business knowledge accessible so agents can perform tasks with it.”
The unspoken corollary: organizations that don’t publish agent-ready knowledge risk becoming invisible to the agents making decisions on behalf of their customers.
The Emerging Practitioner Role
Haynes goes further, sketching the outlines of a new professional discipline. She notes that turning “a company’s mess of data and processes into a clean knowledge graph” is a distinct skill — one that isn’t quite SEO, isn’t quite GEO, and isn’t quite Agentic Search Optimization either. It’s something new.
Her prediction: “SEOs who can make good OKF bundles will be in demand.”
That prediction rests on a specific observation about where the value sits. Building a good OKF bundle requires three overlapping skills:
- Understanding the business deeply enough to identify what knowledge matters
- Decomposing that knowledge into atomic concepts that stand on their own
- Structuring metadata and relationships so agents can traverse the bundle effectively
Why OKF Matters Beyond the Spec Itself
OKF is worth paying attention to not just because of what it is, but because of what it signals. It’s an early, credible attempt to create a shared publishing standard for the agentic era. Whether OKF itself becomes dominant or is superseded by a competing format, the direction is clear:
- Knowledge will increasingly be published in structured, concept-level units
- Those units will be designed for agent consumption first , human consumption second
- The producers who adopt this pattern early will shape how agents perceive their domain
This is the direct link to Graph RAG. As we saw in Section 2, Graph RAG systems spend enormous effort reconstructing structure from unstructured text. OKF removes that burden by shipping structure natively. That connection — the way OKF’s publishing model dovetails with Graph RAG’s retrieval model — is the deeper story we’ll examine in the next section.
What This Means for Generative Engine Optimization (GEO)
If Graph RAG changes how AI systems retrieve knowledge and OKF changes how organizations publish it, the discipline sitting between them — Generative Engine Optimization — has to change too. This section translates the technical shifts into concrete implications for GEO practitioners.
The Atomic Unit of Optimization Is No Longer the Page
For twenty years, SEO has optimized pages. Titles, meta descriptions, H1s, internal links, canonical tags — all of it operates at the page level.
Graph RAG and OKF both point to a different unit: the concept.
- In Graph RAG, retrieval happens at the level of graph nodes and communities, not URLs.
- In OKF, the artifact of publication is the concept file, not the page.
Marie Haynes captures the practical consequence: “a single webpage may end up producing quite a few concept markdown files.” A single product page may decompose into a product concept, a specifications concept, a pricing concept, an integrations concept, and a use-case concept — each individually addressable and individually citable.
GEO implication: Start thinking about your content as a graph of concepts, not a hierarchy of pages. Audit each page and ask: “How many distinct concepts live inside this page, and could each stand on its own as a self-contained answer?”
Structure Beats Prose
Graph RAG pipelines exist to extract structure from unstructured text. That extraction is lossy — ambiguous entity references, buried relationships, and unclear context all degrade the resulting graph.
OKF short-circuits the lossy stages of the pipeline by asking producers to publish structure directly.
GEO implication: The competitive advantage shifts to producers who publish clean structure upfront. That means:
- Explicit entities rather than vague pronouns and references
- Explicit relationships rather than implied connections
- Explicit metadata in schema, frontmatter, or structured data — not just prose
The old SEO advice of “write for humans first” is not wrong, but it’s incomplete. In the agentic era, you write for both — humans read the prose, agents parse the structure.
Visibility Now Requires Two Publishing Layers
Traditional web publishing assumed one output: HTML pages consumed by browsers (and, indirectly, crawlers).
The emerging model requires two:
- The human web — your existing site, optimized for people who visit URLs
- The agent web — structured knowledge (OKF bundles, feeds, APIs, schema-rich data) optimized for AI systems that traverse concepts
These aren’t competing channels. They’re complementary layers of the same brand presence. But they demand different skills, different toolchains, and different quality standards.
GEO implication: Organizations that maintain only the human web are betting their visibility on AI systems being able to reconstruct structure from prose. Organizations that maintain both are giving those systems structure directly — and reducing their dependence on how well the AI’s extraction pipeline happens to work.
Entity Clarity Becomes a Ranking-Adjacent Factor
In Graph RAG’s Stage 2, an LLM extracts entities from your content. If your brand, products, and concepts are described inconsistently — or in ways that could be confused with unrelated entities — they may fail to resolve to distinct nodes in the graph.
This connects to a broader theme visible across the training data: entity attributes matter more than keywords. Systems like Google’s AI Mode operate on entity relationships, not just term frequency. The completeness and consistency of how your brand’s entities are described directly affects whether AI systems can cleanly represent them.
GEO implication:
- Ensure canonical, consistent naming of your brand, products, people, and key concepts across all surfaces
- Build out the attributes AI systems need to disambiguate your entities (categories, relationships, identifiers)
- Use structured data (schema.org) and, where appropriate, OKF-style publishing to make entity boundaries explicit
The New GEO Skill Stack
Marie Haynes predicts that “SEOs who can make good OKF bundles will be in demand.” That prediction implies a specific skill stack:
- Traditional SEO fundamentals — technical crawlability, information architecture, on-page clarity
- Entity and knowledge graph literacy — understanding how AI systems represent and connect entities
- Structured content design — decomposing knowledge into atomic, standalone concepts
- Agent-aware publishing — producing outputs in formats agents can consume natively (OKF, schema, structured feeds)
- Measurement in a probabilistic environment — tracking visibility across AI surfaces where rankings are dynamic and citation is fragmented
Notably, this is not a break from SEO — it’s an extension of it. The practitioners best positioned to move into this space are those who already think in terms of information architecture, entity relationships, and technical structure.
A Strategic Framing: From “Being Found” to “Being Usable”
Haynes offers the clearest strategic framing of the whole shift:
“We will shift from working to be found by search engines to making business knowledge accessible so agents can perform tasks with it.”
That sentence deserves careful reading. The old goal was discovery — making sure a search engine could find and rank your page. The new goal is usability — making sure an agent can retrieve, understand, and act on your knowledge.
Discovery is a means. Usability is the end.
GEO implication: The strategic questions change.
- Old question: “Are we ranking?”
- New question: “When an agent needs to answer a question in our domain, can it actually use our knowledge to do so?”
That reframe alone will reorient most GEO programs. It shifts attention from keywords to concepts, from pages to atomic knowledge units, and from ranking dashboards to agent-readiness audits.
The Convergence: Where Graph RAG and OKF Meet in Practice
So far we’ve treated Graph RAG and OKF as two separate stories — one about how AI systems retrieve knowledge, the other about how organizations publish it. This section brings them together. The core argument is simple: these two developments are not just compatible — they are structurally complementary. One reconstructs; the other pre-structures. Together, they describe a full pipeline from published knowledge to synthesized answer.
Your advantages:
+ Get access to the full exclusive paid articles in the blog.
+ Full analysis of hundreds of well researched active Microsoft, OpenAI and Google patents and research paper.
+ Save a lot of time and get insights in just a few minutes, without having to spend hours analyzing the documents.
+ Get quick exclusive insights about how LLMs, search engines and Google could work with easy to understand summaries and analysis.
+ All patents classified by topic for targeted research.
+ New patent summaries and analysis every week. Weekly notification via E-Mail
+ Get GEO expert knowledge for optimizing your visibility in AI Search via the GEO assistant
+ Use all AI Research Tools to gain insights in seconds from all documents in the patent database, Google Leaks and DOJ trials via the Patent & Paper Analyzer and Google Leak Analyzer
+ Gain fundamental insights for your SEO and GEO work and become a real thought leader.
Get access to the GEO Research Suite and become a thought leader now!
- What is Generative Engine Optimization (GEO)? A new definition and differentiation - 30. July 2026
- Agentic Commerce Optimization (ACO): How Shops Make Products Visible and Orderable in AI Answers and for Agents - 30. July 2026
- Graph RAG Meets the Open Knowledge Format (OKF) — A New Foundation for Generative Engine Optimization - 16. July 2026
- Ultimate Guide to Graph RAG: Why it matters for GEO? - 27. June 2026
- Brand Context Optimization: A Practical Step-by-Step Guide - 26. February 2026
- Brand Identity Blocks for Brand Context Optimization - 25. February 2026
- What is brand context optimization for GEO? - 21. February 2026
- Brand Context Optimization: How to Write Text About Your Brand (for Companies, Persons and Products) - 15. February 2026
- Guide to Brand Context Optimization for Generative Engine Optimization (GEO) - 4. February 2026
- Ultimate guide for llm readability optimization and better chunk relevance - 27. January 2026
