Google's Open Knowledge Format v0.2: The 5 Trust Signals Explained
On July 25, 2026, Google Cloud shipped version 0.2 of the Open Knowledge Format (OKF), adding five trust signals — provenance, trust, freshness, lifecycle and attestation — that let AI agents decide how much to believe a piece of knowledge before they act on it. It landed just six weeks after the first draft, and it has already been swept up in the usual wave of "the next big ranking factor" takes. So here is the honest version: OKF is a real and interesting standard, but it is not a Google Search signal, and treating it like one will waste your time.
This guide walks through what actually changed in v0.2, what each of the five signals does, and — the part most hot takes skip — whether any of it touches your SEO or AI-search visibility. If you have been asked whether your team should "add OKF for GEO," you will have a clear answer by the end.
What is the Open Knowledge Format?
OKF is a vendor-neutral, open specification from Google Cloud, first published as a v0.1 draft on June 12, 2026. In plain terms, it is a way to package organizational knowledge — table schemas, metric definitions, runbooks, business rules — as linked Markdown files with YAML frontmatter, so that AI agents can read it directly instead of guessing from unstructured text.
Google's framing is blunt about the problem it solves: "The context that agents need … should live in a format, not in a proprietary service, and not scattered across unstructured text blobs." An OKF "bundle" is a folder of these Markdown concepts. The only always-required field is type; everything else is optional, and custom keys are preserved rather than rejected. That minimalism is deliberate, and it survives into v0.2.
The key mental model — and the thing that trips up SEO commentary — is who consumes it. OKF lives inside an organization's data and agent stack. As Google puts it, "the most valuable OKF bundles won't be written by hand once and then read forever. They're written continuously, by agents, and consumed by a different set of agents." This is machine-to-machine plumbing, not a public web file like robots.txt.
What changed in v0.2: five trust signals
Version 0.1 could describe knowledge. Version 0.2 lets a consumer judge it. The update adds optional frontmatter that answers five questions an agent should ask before trusting a fact: What was this made from? How much should I trust it? Is it still true? Is it the current version? And was this number produced the way we said it must be?
Crucially, Google chose to record signals, not a credibility score. The reasoning in the announcement is worth quoting: "A score is subjective, doesn't port across consumers" and goes stale fast. Instead, each consumer — an agent, an LLM, an application — reads the raw signals and decides for itself. Here is what each one does.
1. Provenance — "What was this created from?"
A new sources field records the materials a concept derives from: external documents, bundle-relative paths, or scope descriptors. Each source can carry credibility metadata such as author, usage_count and last_modified. The idea is the same instinct a careful human uses — a heavily used, recently updated, authoritatively authored source earns more trust than an anonymous one — expressed as machine-readable fields rather than a gut call.
2. Trust — "How much should I trust it?"
Two fields carry the trust story. generated: { by, at } records how a concept was produced and when it last changed. verified: [ { by, at } ] is a list of independent confirmations — a human sign-off, a machine actor, or several of each. From these, a consumer can derive a simple trust tier: unverified → machine-confirmed → human-reviewed. An agent can then be told, for example, to only act automatically on human-reviewed concepts and to flag the rest.
3. Freshness — "Is it still true?"
A new stale_after field sets a single absolute date after which a concept should be re-checked. Google deliberately picked an absolute date over a relative "time to live," because a fixed date allows a clean, deterministic comparison: is today past stale_after or not? No arithmetic, no ambiguity about when the clock started.
4. Lifecycle — "Is it the current version?"
A status field moves a concept through draft → stable → deprecated. If status is absent, the concept is treated as stable. This lets a bundle keep a deprecated definition around for traceability without an agent mistaking it for current guidance.
5. Attestation — "Was this number produced the way we said it must be?"
This is the biggest structural change in v0.2, and the most novel. A new concept type, Attested Computation, describes a sanctioned way to compute a figure — say, "revenue for a fiscal year" — with declared parameters, an executor, and a receipt. Agents cannot author or edit the computation; they can only run the approved version. A deterministic "attester" then confirms the run matched the sanctioned method. As Google frames it, attestation "confirms a single run produced the value correctly" — it is per-call and never stored in the bundle. The distinction matters: verification is about whether a definition is fresh and reviewed; attestation is about whether a specific number was calculated the right way this time, not improvised with a one-off query.
| Signal | Question it answers | Key field(s) |
|---|---|---|
| Provenance | What was this created from? | sources (author, usage_count, last_modified) |
| Trust | How much should I trust it? | generated, verified |
| Freshness | Is it still true? | stale_after |
| Lifecycle | Is it the current version? | status (draft/stable/deprecated) |
| Attestation | Was this number produced correctly? | Attested Computation type (executor, attester, receipt) |
Every one of these fields is opt-in. A valid v0.1 bundle is still a valid v0.2 bundle; you adopt only the signals you need.
Does OKF affect Google rankings or AI citations?
This is where the marketing conversation runs ahead of the facts, so let's be precise. Neither Google Cloud's announcement nor the OKF specification describes OKF as a web-crawling, ranking, or AI-citation signal. There is no statement — anywhere in the official material — that publishing an /okf/ directory will earn you higher rankings, AI Overview inclusion, or ChatGPT and Perplexity citations.
That is not an oversight; it is the whole design. OKF is built for knowledge that lives inside an organization's agent stack — the metrics and runbooks your internal copilots reason over — not for the open web that Google's crawler indexes. Trust signals like verified and stale_after are read by your agents, or a partner's, to decide whether to act on a fact. They are not read by Googlebot to rank your marketing pages.
If this pattern feels familiar, it should. We saw the exact same hype cycle with another AI-agent-facing format: read our breakdown of what llms.txt is and whether it actually works in 2026, where Google confirmed on the record that Search does not use the file at all. OKF is a more capable standard aimed at a more internal audience, but the SEO lesson is identical — a machine-readable file is only a ranking signal if the ranking system says it reads it, and here it doesn't.
So who should care about OKF?
The honest audience segmentation looks like this:
- Data and platform teams building internal AI agents → This is the real audience. If you are wiring up copilots that query BigQuery, quote metrics, or execute runbooks, OKF's trust signals give you a portable, vendor-neutral way to encode "how confident should the agent be here." Attested Computation in particular is aimed squarely at the "an agent hallucinated a revenue number" problem.
- SaaS and data vendors → Worth watching. If your product ships knowledge or metric definitions that customers' agents consume, emitting OKF bundles could become a genuine interoperability feature — the way exposing an API or a schema is today.
- SEO and content teams chasing AI visibility → This is not your lever. Adding OKF will not get your content cited by AI search, because AI search engines are not the consumers OKF is designed for. Your effort belongs on the signals that AI answer engines actually reward.
For that last group, the fundamentals have not changed. Getting cited by AI Overviews, ChatGPT and Perplexity still comes down to clear, fact-rich, well-structured content and demonstrable expertise. We cover the practical playbook in E-E-A-T for AI search, and the narrower question of markup in whether schema markup helps you get cited. Both will move your visibility far more than any internal knowledge format.
A useful reframe: trust is becoming a first-class field
There is a bigger signal in OKF v0.2 that is worth taking seriously even if you never write a single bundle. The direction of travel across the AI ecosystem is that trust and provenance are turning into structured, machine-readable metadata rather than something inferred loosely from tone or backlinks. Whether it is OKF's verified field for internal agents, or the "who created this and who checked it" instinct behind E-E-A-T for public content, the same question keeps surfacing: can a machine tell how much to trust this, and why?
You cannot "OKF your way" into AI search results. But the underlying idea — make your provenance, freshness, and authorship explicit and verifiable — is exactly what public AI search rewards too. Treat OKF as confirmation of where the whole field is heading, not as a new checkbox for your marketing site.
The takeaway
Open Knowledge Format v0.2, released July 25, 2026, adds five opt-in trust signals — provenance, trust, freshness, lifecycle and attestation — that let AI agents judge knowledge before acting on it, with the standout being the new Attested Computation type for verifying that numbers were calculated the sanctioned way. It is a thoughtful, genuinely useful standard for teams building internal AI agents. It is also, unambiguously, not a Google ranking or AI-citation signal, and no official source claims otherwise. If you run data platforms or agent tooling, dig in. If you run marketing and content, note the direction — trust is becoming structured — and then get back to the fundamentals that actually earn AI citations.
