The open knowledge format (OKF) can tell an AI agent what your content means. It could not tell that agent whether to trust it. Google’s OKF v0.2, released July 25, 2026, adds that layer: five metadata fields covering provenance, authorship, verification, freshness, and lifecycle status. No AI system reads published OKF bundles in production today. But v0.2 reveals, with uncommon clarity, what AI pipelines will eventually require from publishers who want their knowledge to be used.
What is the open knowledge format, and what changed in v0.2?
Google introduced the open knowledge format in June 2026, per the Google Cloud blog. The initial release structured knowledge as linked markdown files with YAML frontmatter (a block of machine-readable metadata at the top of each file), one concept per file. It addressed the structure problem: how do pieces of knowledge connect? Version 0.2 adds a second layer of frontmatter fields with a different job entirely: let a consumer, human or AI agent, make a trust decision before reading a concept’s body at all. According to the Google Cloud blog, “most interactions with a concept never actually progress to accessing the information in the body.” The frontmatter filters. The body is read only after the filter passes.
What five trust questions can OKF v0.2 now answer?
According to the Google Cloud blog (July 2026), every concept in a v0.2 bundle can now carry signals to answer five questions before a consumer reads its body:
- Provenance: what sources did this concept derive from?
- Attribution: who produced this content, and when did it last meaningfully change?
- Attestation: if this concept reports a metric, was it computed the approved way?
- Currency: is this concept still fresh, or has it passed a trust-by date?
- Lifecycle status: is this concept active, deprecated, or in draft?
The new fields carrying these signals: sources (with optional author, usage_count, and last_modified signals), generated (who produced the content and when), and verified (a list of confirmations from human actors, automated processes, or both). All new fields are opt-in. A bundle using none of them remains fully valid under v0.2, per the Google Cloud blog, but “its absence now carries meaning: an unverified concept is distinguishable from a verified one.”
Why did Google deliberately leave out a trust score?
The most significant design decision in OKF v0.2 is what Google chose not to build. Per the Google Cloud blog and Search Engine Journal (both July 2026), OKF records raw signals rather than computing a trust score. The stated rationale: a score “is subjective, doesn’t port across consumers, and goes stale the moment it’s written.” Each consuming system reads the signals and applies its own weighting. From the verified field, three trust tiers emerge:
- Unverified: no verification entry is present.
- Machine-confirmed: confirmation by automated processes only.
- Human-reviewed: at least one human actor signed off.
These tiers are “advisory signals, not access control,” according to the Google Cloud blog.
The fields you cannot fill in honestly are the real audit
Search Engine Journal (July 2026) documents upgrading an existing OKF bundle to v0.2. The finding: adding the fields is fast. Filling them in honestly is not.
The staleness field forces a question that is easy to avoid: how fast does this concept actually age? The SEJ author assigned a 3-month trust window to content about llms.txt, because, as they write, “the story around it changes constantly.” A more stable conceptual framework received a 1-year window. OKF v0.1 allowed all concepts to sit with equal implicit currency. v0.2 requires you to state, on record, which ones go stale.
One field the author could not use: attestation, which is designed for bundles that publish a computed metric alongside a sanctioned formula for recomputing it. That feature is built for data pipelines, not editorial knowledge maps. Recognizing when a field does not apply to your bundle is itself part of the exercise’s value. Search Engine Journal frames the whole process this way: OKF is “a self-audit wearing an export format’s clothes.”
What should you check before your content faces an AI trust filter?
Search Engine Journal (July 2026) suggests a practical starting point: take your most important page and apply the three core questions OKF v0.2 asks of every concept. Before touching the format itself, check:
- Can you name the primary source behind each key claim on that page?
- Do you know the date each claim was last verified by a person?
- Is there a realistic trust-by date: a point after which the claim needs re-checking?
- Was the content generated by a human, an automated process, or both?
- For any metrics or numbers published, can you describe how they were computed?
- Are the relationships between your key concepts documented anywhere outside someone’s head?
What this means for AI-search visibility
AI pipelines are beginning to filter by trust before they filter by relevance. That shift is the practical implication of v0.2, and it matters whether or not a brand ever touches OKF directly.
Picture a user asking an AI assistant about a product. The assistant does not just retrieve the most relevant result. It increasingly checks: who wrote this, did a person verify it, and is it still current? A product page that cannot answer any of those questions may be skipped over, not because it is wrong, but because the agent has no signal to confirm it is right. The competitor whose page carries explicit authorship, a human sign-off, and a clear last-verified date has a structural advantage in that filter pass.
In Hingewise’s assessment, the trust tier distinction has different stakes depending on content type. High-stakes, frequently updated content, such as pricing, product specs, or financial figures, carries real cost when an AI agent acts on stale or unverified data. That category is where human-reviewed verification matters most. Evergreen explanatory content, a definition page or a stable how-to, ages more slowly and may reasonably sit at machine-confirmed. Treating all content identically across tiers is a structural mismatch. OKF v0.2 is, by design, the first open format to make that mismatch visible.
One observation the sources do not raise directly: the no-score design is more durable than it first appears. A trust score produced from within a system carries an inherent tension: you are being asked to accept the system’s judgment of its own output. OKF’s approach, raw signals that any consumer weights for its own context, is closer to how editorial trust has always worked. A journalist’s source list carries more weight than a platform’s auto-generated credibility badge. OKF applies the same logic to structured knowledge, and that logic should age better than any single-number proxy.
OKF v0.2 is not a ranking signal today. It is an early formalization of the questions AI agents will ask when they must choose between your content and a competitor’s. The brands that can answer those questions on record, before that choice happens at scale, will be in a materially stronger position than those that cannot.
Sources: Search Engine Journal, July 2026; Google Cloud Blog, July 2026.
Lam Nguyen · Hingewise