TLDR
Google wants AI assistants to know which information to believe, and it has just published its system for deciding. The idea is simple: a business publishes its facts in a format AI tools can read directly, and each fact carries five plain labels, who it came from, how much weight to give it, how fresh it is, whether it is still current, and whether it has been independently checked. A new receipt mechanism goes further, letting a company prove a number was produced by an approved calculation rather than a best guess. Nothing breaks for early adopters: everything published under the first version still works. The claims come from Google's own documents and no independent audit has been cited.
KEY TAKEAWAYS
What the format does and why Google built it open
The Open Knowledge Format represents organisational knowledge as a directory of UTF-8 markdown files with YAML frontmatter. Each file, called a concept, carries machine-readable metadata alongside human-readable content, so AI agents can discover, filter and consume business knowledge without bespoke APIs or proprietary lock-in.[1]
Sam McVeety, Tech Lead for Data Analytics Engineering at Google Cloud, said the design philosophy was deliberate from the start. "When we introduced the Open Knowledge Format in June 2026, we asserted that the context that agents need, table schemas, metric definitions, runbooks, should live in a format, not in a proprietary service, and not scattered across unstructured text blobs," McVeety said.[1]
The v0.2 update, released on 24 July 2026, builds on that foundation by addressing a gap the first version left open: how an agent should decide whether a piece of knowledge is worth reading before it has read it.[1]
Five signals, one frontmatter block
OKF v0.2 frontmatter carries five trust signals, provenance, trust, freshness, lifecycle and attestation, from which agents can infer source credibility, verification status, currentness and sanctioned computation checks before fetching the full document.verifiedVerified Source: github.com[2] The design logic is economic as much as technical: reading a frontmatter block costs far fewer tokens than parsing a complete markdown file, so surfacing credibility signals early reduces wasted computation across large knowledge corpora.
Amir Hormati, Tech Lead for BigQuery at Google Cloud, said the five fields close a meaningful gap in the original spec. "In OKF v0.2, it is now possible to answer all five of those questions from frontmatter, while the format remains as minimally opinionated as v0.1," Hormati said.[1]
Each signal addresses a different dimension of source quality. Provenance records where a fact originated. Trust encodes a tier or score assigned by whoever published the concept. Freshness sets a window within which the content is considered current. Lifecycle flags whether a concept is active, deprecated or under review. Attestation links the concept to a verification event, which becomes particularly relevant when the underlying value is a calculated figure rather than a manually authored statement.[2]
Attested Computation: proving a number came from the right method
The sharpest addition in v0.2 is a new concept type called Attested Computation. The type enables consumers to verify that a value was produced by an approved method, using a receipt paired with a deterministic attester rather than relying on the publisher's word alone.verifiedVerified Source: github.com[2]
The practical difference matters in analytics-heavy environments. A revenue figure produced by a certified, audited query pipeline carries different weight than a number generated by an analyst running an ad-hoc request against the same data. Without a mechanism to distinguish them, an AI agent consuming that knowledge has no reliable way to know which it is reading. Attested Computation gives that mechanism a standardised representation inside the document itself, so agents and downstream systems can make that distinction programmatically.[2]
The receipt and attester model also creates an audit trail. When an organisation needs to explain why an AI system returned a particular answer, a concept carrying an Attested Computation record can point back to the exact method that produced the underlying value. That traceability is increasingly relevant as regulators in several jurisdictions push for explainability requirements on automated decision-making systems.
Backward compatibility and where to find the spec
The v0.2 update is fully additive and backward-compatible, meaning existing v0.1 bundles remain valid without any modification.verifiedVerified Source: cloud.google.com[1] Teams already working with the format can adopt the new trust signals incrementally, adding them to new concepts or back-filling them into existing ones on their own schedule, without breaking anything already in production.
The full specification, sample bundles and reference implementations are available in the GoogleCloudPlatform/knowledge-catalog repository on GitHub.[1] Publishing the spec there, rather than behind a vendor portal, keeps the format available to any organisation regardless of whether they use Google Cloud infrastructure.
What it means for Australian organisations publishing to AI platforms
The growth of AI-driven search and analytics tools has created a practical problem for Australian businesses: the facts an AI agent surfaces about a brand or product are only as reliable as the knowledge that agent can find and verify. OKF offers one answer to that problem. An organisation that publishes its table schemas, metric definitions and policy runbooks in OKF format, complete with v0.2 trust signals, gives agents a verified factual base to draw on rather than requiring them to infer meaning from unstructured text scattered across a website.[1]
The freshness and lifecycle signals are particularly relevant for organisations in regulated industries, where the currency and status of published guidance can determine whether an agent's answer is appropriate or out of date. A concept flagged as deprecated in its lifecycle field will signal to a compliant agent that it should not be cited in a current context, even if the document remains accessible in the repository.
The attestation signal matters most for data teams whose work feeds downstream AI outputs. A metric definition carrying an attestation record, linking it to a human-verified approval event, sits in a different trust tier from one generated automatically and never reviewed. As AI agents become more common in business intelligence workflows, that distinction will carry real operational weight. Google's v0.2 release does not solve the broader question of AI reliability in those settings, but it gives teams a standardised vocabulary for expressing what they know about the quality of their own data.[2]
Bushletter could not independently verify the technical claims in this article. All facts and figures derive from Google's own release post and the OKF specification, both published on 24 July 2026. No third-party audit or independent review of the format has been cited by Google.
SOURCES & CITATIONS
FREQUENTLY ASKED QUESTIONS
What is the Open Knowledge Format?
What are the five trust signals added in OKF v0.2?
What is Attested Computation?
Do existing OKF v0.1 bundles need to be updated?
Where can I find the OKF v0.2 specification?
Takeshi Mori writes about technology and start-ups. He is curious about how products get built and who they are really for, and he would rather see a thing working than hear it described.
