# Google's Open Knowledge Format: What the v0.2 Trust Signals Mean for AI Search **URL:** https://organikpi.com/blog/geo-ai-search/google-open-knowledge-format-trust-signals/ **Published:** 2026-07-30 **Modified:** 2026-07-30 **Author:** Daniel Shashko > Google Cloud released Open Knowledge Format v0.2 on July 29, 2026, adding five trust signals for agent-maintained knowledge: sources (provenance), generated and verified (trust tiers up to human-reviewed), stale_after (freshness), status (lifecycle), and Attested Computation (verifiable calculations). OKF is an internal format that AI crawlers do not read from websites, so implementing it on a public site has no SEO or citation effect. The practical move for search teams is publishing visible equivalents: per-claim citations, named authors and reviewers, review dates, deprecation notices on outdated posts, and methodology next to statistics. --- > Google Cloud released Open Knowledge Format v0.2 on July 29, 2026, adding five trust signals for agent-maintained knowledge: sources (provenance), generated and verified (trust tiers up to human-reviewed), stale_after (freshness), status (lifecycle), and Attested Computation (verifiable calculations). OKF is an internal format that AI crawlers do not read from websites, so implementing it on a public site has no SEO or citation effect. The practical move for search teams is publishing visible equivalents: per-claim citations, named authors and reviewers, review dates, deprecation notices on outdated posts, and methodology next to statistics. Google Cloud [shipped version 0.2 of the Open Knowledge Format](https://cloud.google.com/blog/products/data-analytics/okf-v0-2-adds-trust-signals) on July 29, 2026, six weeks after the first release. The update adds five trust signals that tell an AI agent where a piece of knowledge came from, who checked it, and when it stops being reliable. The [spec lives on GitHub](https://github.com/GoogleCloudPlatform/knowledge-catalog/blob/main/okf/SPEC.md), it is a folder of markdown files, and you can read the whole thing in twenty minutes. My short version for anyone working in search: you should not implement OKF on your website, and you should still read it. It is the clearest statement so far of how Google thinks machines should decide what to believe. That question is the core of [AI visibility](https://organikpi.com/blog/geo-ai-search/what-is-ai-visibility/) right now, and the five signals map almost one to one onto things you can make visible on your content today. ## What the Open Knowledge Format is OKF is a convention for writing down organizational knowledge so agents can use it: metric definitions, table schemas, playbooks, policies. One concept per markdown file, metadata in YAML frontmatter, files grouped in folders, the whole bundle shipped as a git repository. No SDK, no schema registry, no required tooling. The spec’s own pitch is that if you can run cat on a file you can read OKF. Google Cloud [announced it in June 2026](https://www.searchenginejournal.com/google-cloud-announces-the-open-knowledge-format/579253/) out of its data analytics group, and the examples sit squarely in the BigQuery world: revenue definitions, warehouse schemas, dbt models. The problem it targets is internal sprawl. Knowledge lives in wikis, catalogs, drives, and dashboards, and an agent answering questions from that pile has no way to tell current policy from a doc nobody has touched in three years. The v0.2 spec is blunt about why trust suddenly needed to be first class: a modern corpus is “continuously written and maintained by agents.” When most of your internal docs are machine-written, who says so becomes a required field. From the public repo: a complete v0.2 concept file, the acme_retail revenue computation, with every trust field declared in frontmatter. ## The five trust signals in v0.2 Each OKF v0.2 trust signal answers one question a consumer asks before believing a piece of machine-maintained knowledge. SignalWhere it livesThe question it answersProvenancesources fieldWhat was this created from?Trustgenerated and verified fieldsHow much should I trust it?Freshnessstale_after fieldIs it still true?Lifecyclestatus fieldIs it the current version?AttestationAttested Computation concept typeWas the number produced the sanctioned way? Four details stood out when I read the spec. **There is no credibility score, on purpose.** Each source carries objective signals instead: who authored it, how often it gets used, when it last changed. The consumer decides what those add up to. Google’s reasoning is that a stored score is subjective and goes stale the moment you write it down. Anyone building or buying an AI visibility metric should sit with that for a minute. **Attribution is per claim, not per document.** Version 0.1 had a citations list at the bottom of each file. Version 0.2 retires it. Sources move into frontmatter with stable ids, and individual claims cite them through markdown footnotes keyed to those ids. Google judged a dangling list of links at the end of a document useless for machines. It is just as useless at the end of a blog post. **Verification has tiers, and humans outrank machines.** A concept with no verified field is unverified. Confirmed by an agent or an automated process, it is machine-confirmed. Signed off by an actor with a human: prefix, it is human-reviewed, the top tier. Review by a named person is literally the highest trust state in the format. **Numbers need receipts.** The new Attested Computation type stores the approved way to compute a value, say a revenue query, plus a deterministic checker that confirms the query that actually ran matches the sanctioned one. The spec names the failure mode directly: an agent reporting a dollar figure after improvising its own SQL. From my clone of the repo: the reference viewer rendering the acme_retail bundle. Note the badges on the selected concept: stable, human reviewed, stale after 2026-12-31. ## Why this matters for search people OKF is an internal format, and no AI engine fetches it from public websites. So why write about it here? Because the trust problem it solves is the same one answer engines face when they pick citations. A model deciding which internal metric doc to believe and a model deciding [which page to cite](https://organikpi.com/blog/geo-ai-search/llm-seo/) for a pricing question are running the same evaluation: who wrote this, has anyone verified it, is it current, can the numbers be checked. Google just published, in spec form, the signals it considers first class for that decision. It also extends a pattern I have been tracking all year: infrastructure for agents keeps converging on plain markdown plus explicit metadata. [Agents browsing through MCP servers](https://organikpi.com/blog/technical-seo/ai-agent-browsing-mcp-server-seo/), [Claude Skills](https://organikpi.com/blog/geo-ai-search/claude-skills-content-marketing-ai-citations/), [llms.txt](https://organikpi.com/blog/distribution/llms-txt-adoption-impact/), and now OKF are all bets that machine consumers want small readable files with declared context. OKF is the first of these to make trust metadata the headline feature, and the engines already weigh the same dimensions when they rank, from [freshness weighting](https://organikpi.com/blog/content-strategy/content-freshness-recency-bias/) to [author trust signals](https://organikpi.com/blog/brand-authority/person-schema-author-eeat/). Today they have to scrape those signals off rendered pages. OKF shows what the native version looks like. ## Should you add OKF to your website? No. No search engine or AI crawler reads OKF bundles from public sites. There is no ranking effect, no citation lift, and no crawler that would even request the files. Crawlers fetch your rendered pages, subject to your [crawl permissions](https://organikpi.com/blog/technical-seo/cloudflare-ai-crawl-control-guide/), and that rendered surface is where trust signals get read. [Everything I have measured about AI search this year](https://organikpi.com/blog/geo-ai-search/state-of-ai-search/) points the same way: visible content moves citations, invisible scaffolding does not. If you run internal agent infrastructure, that is a different conversation. A knowledge base of metric definitions with verified and stale_after fields would have saved several teams I have worked with from agents confidently quoting dead numbers. But that is a data platform decision, unrelated to your public site. ## What to copy: make the five signals visible The useful move is translating each signal into something a reader and a crawler can both see on your pages. This is E-E-A-T mechanics with a spec behind it. - **Provenance.** Cite the source next to the claim it supports, with the source named in the sentence. A references dump at the end of the post is the exact pattern Google just deprecated in its own format. - **Trust.** Put a named author with real credentials on the page, and when someone checks a piece, say who and when. Keep your [entity footprint](https://organikpi.com/blog/technical-seo/wikidata-entity-seo-knowledge-graph/) consistent so machines can resolve who that person is. - **Freshness.** Show an honest updated date and an actual review cadence. “Reviewed July 2026, next review Q4” is a stale_after field a human can read. - **Lifecycle.** When a post is superseded, mark it and point to the current version instead of quietly deleting it. Deprecated content that declares itself deprecated keeps its links and loses its ability to mislead. - **Attestation.** Publish your method next to your numbers: sample size, query set, date range, tooling. If a model can recompute your statistic, it can trust your statistic. Structure it in [clean semantic HTML](https://organikpi.com/blog/technical-seo/html5-semantic-tags-ai-content-extraction/) so extraction is trivial. If you are earlier in the journey, start with [what GEO actually is](https://organikpi.com/blog/geo-ai-search/what-is-geo-generative-engine-optimization/), then audit which of the five questions your most important pages leave unanswered. The agents are already asking them. ## Frequently Asked Questions ### What is Google's Open Knowledge Format? OKF is an open specification from Google Cloud for recording organizational knowledge as markdown files with YAML frontmatter so both humans and AI agents can read it. A bundle is a folder of concept files, usually shipped as a git repository, with no SDK or schema registry required. It launched in June 2026 and version 0.2 arrived on July 29, 2026. ### What did OKF version 0.2 add? Five trust signal families: sources for provenance, generated and verified for authorship and review, stale_after for freshness, status for lifecycle, and a new Attested Computation concept type that lets a consumer confirm a value was produced by the sanctioned computation rather than SQL an agent improvised. ### Does OKF affect SEO or AI search rankings? No. OKF is an internal knowledge format for agent infrastructure. No search engine or AI crawler fetches OKF bundles from public websites, so implementing it on a site has no ranking or citation effect. The trust concepts behind it, per-claim sources, named review, and freshness dates, are worth applying to visible page content. ### Should I implement OKF on my website? Not for visibility. Publish the human-readable equivalents instead: citations next to each claim, a named author and reviewer, an honest updated date with a review cadence, deprecation notices on superseded posts, and your methodology next to any statistic you want cited.