Google's OKF Is the Feed Layer. LLM Search Console Is the Measurement Layer. Here's Why Both Matter.
Google released the Open Knowledge Format on June 12, 2026. Version 0.1. Quiet announcement. A GitHub repo. A spec that fits on one page.
Most marketers either missed it or filed it under “things to look at later.” That’s a mistake — not because OKF is a ranking signal (it isn’t), but because it operates at exactly the layer that determines your brand’s accuracy inside AI-generated answers. And if you’ve been using LLM Search Console to track your AI Share of Voice, you already know that accuracy is the new ranking.
Let me connect the dots between OKF and what we measure every day inside LLM Search Console.
Thanks for reading Codedesign Agency - Your Digital Transformation Partner ! Subscribe for free to receive new posts and support my work.
The Problem LLM Search Console Keeps Surfacing
When we built LLM Search Console, the core premise was simple: Google Search Console tells you how bots see your site. My tool tells you how AI models see your brand. What ChatGPT says when someone asks about you. What Gemini cites. What Perplexity surfaces. Whether your brand appears at all in competitive queries — and when it does, whether what the AI says is accurate.
That last part is where things get interesting.
We monitor hundreds of prompts across ChatGPT, Claude, Gemini, and Perplexity for clients at Codedesign. And one of the most consistent findings is this: brands don’t just suffer from not being mentioned in AI answers. They suffer from being mentioned incorrectly. Wrong pricing tier. Outdated product description. A use case framing that no longer reflects their positioning. An AI confidently describing a feature that was deprecated two years ago.
This isn’t a crawling problem. It’s a knowledge structure problem. And OKF is, in many ways, Google’s formal answer to it.
What OKF Actually Is — And Why It’s a GEO Signal
OKF is a vendor-neutral, open specification for structuring organizational knowledge as a directory of Markdown files with YAML frontmatter. Each file represents one concept — a metric, a product definition, a process, a dataset. The only required field is type. Cross-links between concept files turn the directory into a knowledge graph. The full spec fits on a single page.
Here’s a minimal OKF concept file:
---
type: service
title: Enterprise Plan
description: Full-access tier for teams of 10+, including custom integrations and a dedicated LLM search
tags: [pricing, enterprise, tier]
---
# Enterprise Plan
The Enterprise Plan includes unlimited prompt monitoring, custom competitor sets, white-label reporting, and a dedicated Customer Success Manager. Billed annually.
Related: [Pro Plan](./pro-plan.md), [Pricing FAQ](./pricing-faq.md)
That’s it. One concept. Machine-readable without a bespoke SDK. Diffable in Git. Human-readable without tooling.
The GEO angle: when an AI agent reads your OKF bundle before generating an answer about your brand, it’s grounding on a single canonical source rather than stitching together fragments from your homepage, a TechCrunch article from 2023, and a Reddit thread where someone complained about your old pricing. The result isn’t just more accurate — it’s more citable. Structured, attributed knowledge is exactly what citation analysis tools like LLM Search Console track as the upstream driver of AI Share of Voice.
OKF is the feed layer. LLM Search Console is the measurement layer.
The Karpathy Connection: This Has Been Building
A few months ago, Andrej Karpathy posted a short GitHub gist outlining what he called the LLM Wiki pattern. The idea: stop treating AI like a search engine where every query starts from zero. Instead, let the LLM compile your knowledge into a structured wiki first — then answer questions from that compiled artifact.
His analogy was from software engineering: source code doesn’t run directly. It gets compiled into an optimized binary and executed efficiently on demand. The compilation step is expensive once. The runtime benefit compounds across every query afterward.
OKF is the organizational-scale, open-spec formalization of that pattern. Karpathy was describing a personal productivity workflow. Google is describing an enterprise knowledge infrastructure standard. Same underlying insight: AI works better when knowledge is compiled once, structured canonically, and served consistently — rather than re-derived from scattered, contradictory sources every single time.
For LLM Search Console users, this maps directly to something we’ve been tracking in citation analysis: pages that score highest for “crit-ability” — the likelihood of being cited in an AI-generated response — tend to be single-concept, densely structured, and cross-linked. OKF concept files are precisely that architecture at scale.
The Context Tax Is Real — And OKF Addresses It
I wrote about this in an earlier piece here: context windows are your real GEO bottleneck. The obsession with model size misses the point. What matters isn’t how big your context window is — it’s how much of it you’re actually using for reasoning versus burning on redundant grounding.
OKF reduces the context tax. Here’s how.
When an AI agent needs to answer a question about your brand without a structured knowledge source, it has to do expensive retrieval — RAG chunks from multiple documents, partial embeddings, stitched together under time and token pressure. That process is noisy, redundant, and eats into the context window before reasoning even begins. The result: more hallucinations, lower accuracy, lower citation confidence.
A well-built OKF bundle changes the math. Each concept file is already compressed, already structured, already cross-linked. The agent can find your Enterprise Plan definition in one file, follow the link to your Pricing FAQ, and generate a precise, well-grounded answer in far fewer tokens. Less context tax. More reasoning space. Higher accuracy in the output.
In LLM Search Console terms: lower perplexity on your brand-specific queries. Better factual consistency across models. Higher AI Share of Voice in competitive prompts.
What I’ve Seen in Client Data
One of our clients — a B2B software company selling supply chain tooling to enterprise buyers in Northern Europe — noticed something odd in their LLM Search Console prompt monitoring reports. When we tracked the prompt “What are the best [category] tools for enterprise teams?” across ChatGPT, Gemini, and Perplexity, the client appeared in roughly 30% of responses. Decent visibility. But when we ran citation analysis on how they appeared, something stood out: two different LLMs were describing their product with conflicting feature sets. One said they offered native ERP integrations. One said they didn’t.
Both were technically pulling from real content — but from different points in time. An old help article pre-dating a product update, and a newer landing page that had been rewritten. No single canonical source existed.
We helped them build what is now essentially an YAML bundle: a structured Markdown directory of product definitions, integration specs, and pricing tier descriptions — all version-controlled, all cross-linked, all accessible to their internal agents and (via their llms.txt routing) to external crawlers.
We reran the same prompts four weeks later. Citation consistency across models improved significantly. Both models now cited the same integration capability. One of them added an unprompted positive comparison to a competitor.
They didn’t change their content. They changed how their knowledge was structured. That’s the OKF thesis applied in the real world — and LLM Search Console is exactly the tool you use to measure whether it’s working.
OKF, llms.txt, and the Three Layers of AI Visibility
I want to be precise here because conflation of these tools is rampant in GEO commentary right now.
Layer 1 — Discovery (llms.txt): A public file at your domain root that tells external AI crawlers what your most important pages are. This is the routing layer. It doesn’t affect Google rankings. It tells agents where to go.
Layer 2 — Structure (OKF): An internal or semi-public knowledge bundle in structured Markdown. This is the grounding layer. It tells agents what your organization knows, defined precisely, in a machine-readable format that compresses cleanly into context windows.
Layer 3 — Measurement (LLM Search Console): Continuous monitoring of how AI models actually perceive your brand across prompts, competitors, and citations. This is the feedback layer. It closes the loop — tells you whether Layer 1 and Layer 2 are working.
You can do Layer 1 without Layer 2. Most brands with a llms.txt file have done exactly that: they’ve told agents where to look, but the pages agents land on are still structured for human persuasion, not machine comprehension. OKF is the missing piece that makes Layer 1 actually effective.
And without Layer 3, you’re flying blind. You don’t know if your efforts are changing AI perception. You don’t know if a competitor is gaining Share of Voice on a key query. You don’t know if an LLM is still citing your old pricing because it grabbed a cached snippet before your OKF bundle existed.
All three layers work together. If you’re only doing one or two of them, you’re leaving measurable ground on the table.
How to Build Your First OKF Bundle — With GEO in Mind
It’s all taking the first steps… Start with the concepts that LLMs most often get wrong about your brand. Your LLM Search Console citation analysis will tell you exactly what those are: look at the prompts with the lowest factual consistency scores and work backward to the underlying knowledge gaps.
Step 1: Run a brand accuracy audit. In LLM Search Console, pull your brand-relevant prompts and filter for responses with factual inconsistencies across models. These inconsistencies reveal your knowledge gaps.
Step 2: List the 10-15 concepts that need canonical definitions. Your main product tiers, key integrations, competitive differentiators, company facts (founded, HQ, team size), use case categories. These become your first OKF concept files.
Step 3: Write one Markdown file per concept. Keep it tight. The required field is type. Add title, description, and tags. Write the body as a single-concept explanation that would read as an accurate, authoritative answer to a query about that concept. Cross-link related concepts.
Step 4: Add an index.md. Think of it as the agent’s onboarding document. “Start here. The most important product concepts are X, Y, Z. For competitive comparisons, see [comparison.md].”
Step 5: Version it and expose it. Commit to Git. If you want external agents to read it, reference the bundle location in your llms.txt. If it’s internal, point your AI tooling at the directory as its primary grounding source before any web retrieval.
Step 6: Re-run your LLM Search Console prompts. Measure factual consistency before and after. Watch your citation quality scores. This is your feedback loop.
The Agentic Web Is the New Search Engine
Everything I’ve been tracking with LLM Search Console points to the same macro shift: the web is being consumed increasingly by agents, not humans. Sales-research agents operating inside CRMs are summarizing your product for buyers who never visit your site. Procurement tools are using AI to shortlist vendors based on capability queries. Financial analysts are using AI to compare competitive positioning.
In that world, the quality of your structured knowledge is a direct determinant of how well you compete. Not your PageRank. Not your Domain Authority. How accurately and confidently an AI agent can describe what you do, why you’re different, and why you’re the right choice — based on the knowledge you’ve made available in a format agents can actually use.
OKF is an early, imperfect spec. Google calls it “a starting point, not a finished standard.” But the pattern it formalizes — structured, canonical, agent-readable knowledge — is the direction the entire GEO discipline is moving toward.
The organizations that build this infrastructure now will have compounding advantages as AI agents become the primary interface between buyers and brands. And the organizations that track whether it’s working — using tools built specifically for AI visibility, not tools repurposed from traditional SEO — will be the ones who actually close the loop.
What Are You Seeing in Your Data?
If you’re an LLM Search Console user, I’d genuinely love to know: are you seeing factual inconsistencies in how different models describe your brand? Have you experimented with llms.txt, structured content, or anything resembling OKF in your knowledge architecture?
Drop a comment below or reach out directly. The community building around GEO is still early — the people sharing data and real-world results are the ones setting the direction for everyone else.
And if you’re not yet tracking your AI Share of Voice, start here. The measurement layer is where the strategy starts.
Bruno Gavino is the Founder of LLM Search Console and CEO of Codedesign. He also publishes the Voice of Experts series on AI, marketing, and the agentic web.