OKF vs llms.txt: which one actually reaches AI assistants?

Within a week of Google Cloud publishing the Open Knowledge Format, I started seeing it filed next to llms.txt as the next thing a brand should add to its site. I understand the hype, and I understand wanting to be early, but this article will show you where that bundle really sits and what your expectations should be. The two were built for different readers, and neither of them is the reader that matters when someone asks ChatGPT or Perplexity about your company, while the page a crawler fetches is what reaches the assistant, which is why OKF is still the more interesting of the two documents to read.

What OKF is, and who it was written for

Google Cloud published the Open Knowledge Format on June 12, 2026, as a standalone, Apache-licensed specification, now hosted in the GoogleCloudPlatform/open-knowledge-format repository (Google Cloud, June 2026). A bundle is a folder of markdown files, each one carrying a YAML block at the top, and in version 0.1 the only mandatory key was type, with title, description, resource, tags and timestamp optional, plus an index.md listing the contents and an optional log.md for changes. Google shipped a BigQuery enrichment agent, a static visualizer and three sample bundles alongside it, covering GA4, Stack Overflow and Bitcoin data, and the announcement credits Andrej Karpathy's "LLM wiki" idea as the pattern it formalizes.

You could take a glance at those sample bundles and you would understand how and where Google will use this information. They describe tables and metrics for enterprise data teams, and Google updated Knowledge Catalog, the data-governance product it renamed from Dataplex Universal Catalog on April 10, 2026, to ingest OKF bundles (Dataplex release notes). Some coverage implied OKF ships inside that product, while the spec is a standalone Apache-licensed repository and Knowledge Catalog was simply taught to read it, and that detail tells you where Google expects bundles to live, which is inside a company's own data platform, behind its own access controls.

Is OKF a ranking factor?

To be honest, no. The specification comes from Google Cloud, it has no stated connection to rankings, AI Overviews or AI Mode, and the SEO press that looked at it closely came back with the same conclusion. I believe the confusion comes from the timing, because OKF arrived in the same summer that marketing teams were being told to add llms.txt, and a second markdown-for-agents format looked like a second box to tick.

The more useful test is to ask who fetches the file. As of this writing, OpenAI, Anthropic, Perplexity, Google Search, Cloudflare and Vercel have not announced any crawler or assistant that looks for an OKF bundle on a public site, and the reference tools Google ships all assume the bundle is sitting in a place the agent already has access to. No matter how well the format is designed, an assistant that never fetches it cannot repeat what is in it.

What llms.txt delivers in practice

llms.txt at least sits at a public path that a visiting agent could read, so it deserves a fairer hearing than OKF on the visibility question. The concession I will make is that it is cheap, harmless and occasionally fetched. The evidence that it changes anything is thin, and the numbers are recent enough that I will state them with their dates attached.

Casey Burridge's June 2026 pass over HTTP Archive data found llms.txt on 421 of 7,504 crawled sites from the top 10,000, about 5.61 percent (Casey Burridge, State of llms.txt adoption), and Rankability's Tranco sample put it at 8.7 percent of the top 1,000 in the same month (Rankability, llms.txt adoption data), which is fast growth from a standing start and still a small minority. The more sobering figure is on the demand side, where Ahrefs checked 137,210 domains in May 2026 and found that about 97 percent of published llms.txt files received zero requests of any kind (Ahrefs, llms.txt study, 2026). Search Engine Land ran a ten-site test in January 2026 and found no change in AI referral traffic on eight of the ten sites after adding the file (Search Engine Land, Does llms.txt matter?). Ten sites is a small sample, and I would not treat it as the final word, but it lines up with the request data, which is that most of these files are written for a reader who does not show up.

So llms.txt is a signpost. It tells an agent which pages you would like read, and a signpost is only useful when someone walks past it and then follows it, which the logs suggest is rare.

The layer that does reach assistants

What the assistants read is less mysterious than the format debate makes it sound. GPTBot, ClaudeBot, PerplexityBot and Google's crawlers request ordinary URLs, they receive whatever HTML the server returns, and the model works from that text plus the structured data embedded in it. When Perplexity lists a competitor's pricing under your name, the model most likely fetched your pricing page, found the numbers buried in a layout it could not read, and went with a third-party page that stated them plain and simple. A signpost file would have changed very little about that fetch. I wrote about how that gap shows up in practice in why ChatGPT does not mention your company, and the mechanics of a generated answer versus a ranked link are in GEO is the new SEO.

Put the three side by side and you can see that each one was written for a different reader.

The third column is the one Ooky works on. Brand Intelligence serves eligible AI crawlers an AI-ready version of each page at the edge, with the facts you have confirmed written out in full sentences, and a JSON-LD block alongside them, so a model that fetches your pricing page finds your pricing stated in plain sentences. That is the layer where the evidence points, and I would rather spend a brand's effort there than on a file whose reader may never arrive.

Why I still read OKF v0.2 closely

Having said all of that, I believe OKF is the more interesting document of the two, and the reason arrived six weeks after launch. On July 24, 2026, Google Cloud published version 0.2 (Google Cloud, OKF v0.2 adds trust signals), and the additions are all about accountability. A document can now carry a sources list, generated and verified entries where a human reviewer is marked with a human: actor prefix, which yields three visible trust tiers of unverified, machine-confirmed and human-reviewed, a status of draft, stable or deprecated, a stale_after date after which the claim should not be trusted, and Attested Computation fields naming an attester and carrying an executor receipt for a number that came from a sanctioned calculation. The release also renamed timestamp to generated.at and moved body-level citations into sources, while keeping the old forms working.

Schema.org has carried reviewedBy and expires for years, so the idea is older than OKF, and still I think this is the first time a major vendor has made a named reviewer and an expiry date the center of a knowledge format. You get something out of that discipline even if no agent ever opens the bundle. One practitioner who bundled his own site found that of 61 concepts, only 18 had a verified entry and 8 were already past their stale_after date (Suganthan, Open Knowledge Format). The format audited his content simply by asking him to fill in the fields, and most marketing sites would fail the same audit, because nobody has ever been asked to write down who checked the integrations page or when the pricing stops being true.

That discipline is the same one behind the pages Ooky publishes. Every fact in a Brand Intelligence page is checked against the visible page it came from before it goes live, the page carries a publish status so a stale claim is visible as stale, and for the fields a model must never guess at, a do_not_infer directive tells it to leave the answer blank instead of inventing a tier. OKF put a Google-shaped container around that idea for internal agents, and I take that as confirmation that the industry is converging on the same rule, which is that an unverified claim with no expiry is a liability wherever it lives.

If you want to see how your own brand reads to the crawlers today, the free tier tracks what Gemini and Perplexity say about you on a schedule, so you can see the gap before you decide whether to pay for the publishing layer.

FAQ

Keep reading

• Decision·August 28, 2026·9 min

  What AI visibility platforms actually cost in 2026

  Read the post →

• Playbook·August 24, 2026·8 min

  How to read your AI crawler logs, and what GPTBot is telling you

  Read the post →

Sources

1. Google Cloud. "How the Open Knowledge Format can improve data sharing." June 12, 2026. Launch announcement, v0.1 field list, reference tools and sample bundles. Retrieved 2026-08-24. https://cloud.google.com/blog/products/data-analytics/how-the-open-knowledge-format-can-improve-data-sharing
2. Google Cloud. "OKF v0.2 adds trust signals." July 24, 2026. Sources, generated and verified entries, status, stale_after and Attested Computation fields. Retrieved 2026-08-24. https://cloud.google.com/blog/products/data-analytics/okf-v0-2-adds-trust-signals
3. Google Cloud Platform. "open-knowledge-format" repository, Open Knowledge Format specification, Apache 2.0 license. Retrieved 2026-08-24. https://github.com/GoogleCloudPlatform/open-knowledge-format
4. Google Cloud. Dataplex release notes, "Dataplex Universal Catalog is now called Knowledge Catalog." April 10, 2026. Retrieved 2026-08-24. https://docs.cloud.google.com/dataplex/docs/release-notes
5. Search Engine Journal. "Google Cloud announces the Open Knowledge Format." June 2026. Retrieved 2026-08-24. https://www.searchenginejournal.com/google-cloud-announces-the-open-knowledge-format/579253/
6. MarkTechPost. "Google Cloud introduces Open Knowledge Format (OKF), a vendor-neutral markdown spec for giving AI agents curated context." June 16, 2026. Retrieved 2026-08-24. https://www.marktechpost.com/2026/06/16/google-cloud-introduces-open-knowledge-format-okf-a-vendor-neutral-markdown-spec-for-giving-ai-agents-curated-context/
7. Suganthan. "Open Knowledge Format: Google's new markdown format for AI agents." 2026. Practitioner bundle of 61 concepts, 18 with a verified entry, 8 past stale_after. Retrieved 2026-08-24. https://suganthan.com/blog/open-knowledge-format/
8. Casey Burridge. "State of llms.txt adoption." HTTP Archive data, June 2026, 421 of 7,504 crawled top-10,000 sites. Retrieved 2026-08-24. https://caseyrb.com/blog/state-of-llms-txt-adoption/
9. Rankability. "llms.txt adoption." Tranco top 1,000 sample, June 2026, 8.7 percent adoption. Retrieved 2026-08-24. https://www.rankability.com/data/llms-txt-adoption/
10. Ahrefs. "llms.txt study." 137,210 domains, May 2026, about 97 percent of llms.txt files received zero requests. Retrieved 2026-08-24. https://ahrefs.com/blog/llmstxt-study/
11. Search Engine Land. "Does llms.txt matter?" January 20, 2026. Ten-site, 90-day test of AI referral traffic and crawl frequency. Retrieved 2026-08-24. https://searchengineland.com/does-llms-txt-matter-467740