Google Just Defined AI Trust: 5 Signals, Zero Score

News: On 25 July 2026, six weeks after the first release, Google Cloud shipped OKF v0.2, an update to the Open Knowledge Format that adds five families of trust signal: provenance, verification, freshness, lifecycle and attestation. The question the announcement opens on is blunt: "Once agents are writing to the corpus, can it really be trusted?" (Google Cloud Blog, Sam McVeety and Amir Hormati, 25 July 2026).

Direct answer: OKF v0.2 is an infrastructure specification aimed at enterprise agents, not the public web, and it has no effect on your Google rankings. Its value to a publisher lies elsewhere. Google engineers have just written down, in a machine-readable format, the list of what an AI system wants to know before reusing a piece of information. There are five entries. Where it came from, who produced it and who verified it, how long it stays valid, where it sits in its lifecycle, and whether the number inside it can be recomputed. None of those five questions is about word count, keyword density or internal linking. All of them are about traceability.

What v0.2 actually adds

Quick recap for anyone who missed it: OKF is an open specification published on 12 June 2026 by Google Cloud's Data Cloud teams. It represents a knowledge base as a plain directory of Markdown files with YAML frontmatter, versionable in git, readable by a human and by an agent alike. We covered the format and its limits in our analysis of v0.1 back in June.

The problem v0.2 solves came out of usage: once agents are generating thousands of knowledge files themselves, another agent reading them has no way to tell whether what it is reading was verified, by whom, and whether it is still true. v0.2 answers by adding frontmatter vocabulary, and nothing else: per the announcement, it "adds vocabulary, not rules", type remains the only always-required field, and every new field is opt-in.

Two design details are worth flagging. The stale_after field takes an absolute date rather than a duration: the authors say they chose that form deliberately so staleness becomes "a plain date comparison with no reference to when the concept happened to be read". And the trust ladder derived from the verified field has only three rungs: no verified key means unverified, verification by non-human actors means machine-confirmed, and verification by a human:<id> actor means human-reviewed. The top rung is human by name.

The most interesting part: there is no score

This is the line to take away from the whole announcement, and it runs to eight words: OKF records the signals, not a credibility score. The authors add that credibility is inferred from the signals by whoever is consuming the information, and can be dynamically scored by that consumer if desired.

Transpose that architecture onto your content. You get an accurate description of what happens when a generative engine decides to cite you, or skip you. There is no authority grade filed away next to your domain name. There never was. There is a system that, at answer time, looks for traceability markers in what is in front of it and draws a conclusion. A page with no named source, no author and no date is not lowering a score: it is giving nothing to compute. That is the mechanism we observed when analysing why commodity content never gets cited by AI, and it also explains why pages outside Google's top 10 get cited while better-ranked pages do not.

The five signals, translated for a website

OKF is not built for the public web and trying to deploy it there would be pointless. The five questions it formalises, however, translate directly, using tools that already exist.

• Provenance. Cite named, dated, linked sources rather than writing that "studies show". A claim whose origin can be checked is reusable; the rest is noise.
• Verification. Show an identifiable author, their competence on the subject, and a last-verified date distinct from the publication date. That is also, near enough, the editorial exemption condition the EU AI Act's Article 50 makes enforceable on 2 August. The regulator and the technical spec are asking for the same proof.
• Freshness. State explicitly how long a data point stays valid, instead of letting a 2023 figure float unlabelled. A price, a regulatory threshold or a market statistic has a known expiry date: write it down.
• Lifecycle. Mark obsolete pages instead of leaving them online as-is. A deprecated page that says so and points to the current version beats a dead page that gets cited by mistake.
• Attestation. Publish the method behind your proprietary numbers: sample, period, how it was computed. That is what turns an in-house data point into a citable one, and it is the only one of the five signals your competitors cannot copy.

What to do this week

1. Take your ten most strategic pages and check, for each, whether all five questions have a visible answer on the page. Not in your CMS: on the page.
2. Separate publication date from verification date. Plenty of sites show only the former, or worse, rewrite the former to look fresh. That is the opposite of the signal being sought.
3. Date your figures inside the copy, not just in the footer. An extract quoted by an AI rarely carries the page's context with it.
4. Deal with the obsolete backlog instead of producing more. An unflagged deprecated page is an error source you are feeding yourself. It is one of the checks in our GEO audit method.

Our take

v0.1 said how to file knowledge. v0.2 says on what terms we agree to believe it, and the answer fits in one sentence: show where it came from, who checked it, and how long it holds. That this vocabulary was born for BigQuery tables rather than blog posts changes nothing about what it reveals. When the engineers building the reading machines write down their own definition of trust, it is worth reading.

What this article does not cover

OKF is a Google Cloud specification for enterprise knowledge systems. Neither the 25 July announcement nor the specification connects it to Google Search, AI Overviews or web page ranking: the transposition to editorial content proposed here is our reading, not a Google recommendation, and it should not be presented to a client as a ranking factor. We also have no adoption data: as of publication, no public figure shows how many organisations use the format. Finally, v0.2 introduces two breaking changes against v0.1. timestamp was renamed to generated.at, and body citations moved into the sources frontmatter. If you had started experimenting, check those two points. We will update this article if Google publishes adoption data or a v0.3.

Frequently asked questions

Alexis Dollé

CEO & Founder

Growth and SEO content strategist, I founded Cicéro to help businesses build lasting organic visibility, on Google and in AI-generated answers alike. Every piece of content we produce is designed to convert, not just to exist.

LinkedIn