AI Web Layers Are Multiplying: What Operators Must Know

TL;DR: Google shipped two new machine-readable web specs — Open Knowledge Format (OKF) and Agentic Resource Discovery (ARD) — within days of each other in mid-2026. They sit at different layers of the emerging agentic web and serve completely different purposes. For most high-spend operators in regulated verticals, the practical answer right now is: understand the stack, don’t panic, don’t rebuild.

The Web Is Not One Thing Anymore

The web operators run campaigns on is a crawlable, HTML-based, link-structured environment. That is not going away. What is happening is that a parallel machine-readable layer is being built on top of it — and that layer is not a single thing either. It is a stack of at least six distinct formats, each doing a different job:

• Crawlable HTML — still the foundation; nothing here has changed.
• Schema.org / Structured Data — semantic vocabulary baked into pages to tell machines what content means.
• LLMs.txt — a navigation file for AI agents already on your site, pointing them toward the pages that matter.
• MCP / WebMCP — standardized protocols for AI to connect to services and take action, beyond individual pages.
• Open Knowledge Format (OKF) — a directory of markdown files with YAML headers, designed for packaging structured internal knowledge.
• Agentic Resource Discovery (ARD) — a new open spec for how agents find and verify tools, skills, and other agents across the web at runtime.

Each layer does something different. The mistake most operators are making — and most LinkedIn hot-takes are reinforcing — is treating all of it as a single “AI SEO” initiative. It is not. Running a full-stack marketing audit right now should include identifying which of these layers, if any, are relevant to your specific site architecture and business model.

What OKF Actually Does (And What It Does Not)

Google published the OKF spec quietly, attached to a rebrand of Dataplex into Knowledge Catalog. The format itself is straightforward: a directory of markdown files, each with a YAML header declaring a type, title, description, resource URL, and tags. The files link to each other. That’s the whole thing.

OKF is not a retrieval system. It does not replace HTML crawling. It does not magically surface your site as the preferred source inside an LLM’s retrieval pipeline. Any self-reported format gets gamed eventually — structured data history proves this. FAQ schema had a strong run in SERPs until Google discontinued the rich result. The platforms absorb the signal, fold it into the algorithm, and the explicit markup becomes less decisive over time. OKF will follow the same arc.

The more important context: OKF was built for data teams sharing internal knowledge — table schemas, runbooks, metric definitions — between AI agents inside organizations. Applying it to a public marketing site is a stretch. It adds one higher-signal input into an already complex retrieval pipeline. It does not replace that pipeline. For operators running iGaming acquisition programs or similar content-heavy verticals, OKF may eventually be worth testing on documentation and product spec pages. It is not worth restructuring your site for.

The Markdown Conversion Question: Still No

Every few weeks a new argument surfaces for converting your site to markdown to improve AI readability. The answer remains no, and the reason is structural, not aesthetic.

HTML carries information that markdown strips out: navigation hierarchies, internal link relationships, header structures, footer context. These are not decoration — they are how crawlers understand your site’s shape. Removing them to marginally improve per-page machine readability breaks the discovery mechanisms that got your pages indexed in the first place.

As SEO analyst Jono Alderson put it precisely: a page is not just a container for words. It is an editorial artifact. Hierarchy, emphasis, what comes first, what sits in a footnote — these are signals about meaning. When you flatten a page to markdown, you remove judgment and context. You also create a second version of your content that now needs to be maintained separately, which is exactly the parallel-rendering problem the SEO industry already solved the hard way with JavaScript rendering.

Google’s John Mueller confirmed this on Search Off the Record: standard HTML pages remain critical for crawlers and LLM systems alike, because structural information is how they understand your site’s architecture. The one legitimate exception is developer documentation, where a markdown file genuinely helps an agent parse an API reference. Outside that use case, stay with semantic HTML.

For operators running crypto acquisition funnels or any program where landing page architecture is a conversion variable, the risk of markdown conversion is not theoretical — you are actively degrading the signals that drive indexability and link equity.

What ARD Is Actually Building

Agentic Resource Discovery is a different problem entirely from OKF. It was announced by Google on June 17, 2026, and the distinction matters: OKF is about packaging knowledge for consumption; ARD is about advertising capabilities for connection.

The problem ARD solves is coordination at scale. Right now, an AI agent has to be pre-configured to know about every tool, MCP server, or API it will use. That works at small scale. It stops working when the number of available capabilities grows faster than any team can pre-configure manually. ARD moves discovery from setup time to runtime. The agent finds what it needs when it needs it.

The spec is built on two components: Catalogs (an ai-catalog.json file hosted on your domain describing available capabilities) and Registries (effectively search engines for the agentic web, crawling catalogs and returning matching capabilities with verified publisher metadata). Hugging Face has already integrated ARD into its Discover Tool.

The identity and trust layer underneath ARD is still being settled. There is also an unresolved gap: no agreed media type exists for an OKF bundle yet, which means autonomous agents ingesting catalog entries may act on data they cannot properly verify. For operators considering building AI-powered lead qualification agents or programmatic service interfaces, ARD is the direction the infrastructure is heading — but it is infrastructure to monitor closely, not build critical business paths on yet.

What This Means for High-CAC Vertical Operators

Forex, legal, crypto, and iGaming operators spend $10,000 to $100,000-plus per month acquiring leads in regulated, competitive environments. The question is never “is this new format interesting?” It is “does this move my cost-per-acquisition or my lead quality?” Right now, the honest answer on OKF and ARD is: not directly, not yet.

What does move CAC in these verticals is the same stack it always has been: authoritative content that earns links, a brand presence that users actually recognize and search for, structured data properly implemented where it drives rich results, and paid acquisition tightly managed against real attribution data. Operators running legal marketing programs do not need to ship OKF bundles before they have clean site architecture and a defensible content authority position.

For forex lead generation, the same logic applies: broker sites where half the pages have thin content and weak internal linking are not going to recover by adding YAML frontmatter. Fix the foundation first. Similarly, for CDL recruitment marketing, where the conversion path runs through job-board integrations and local search signals, the agentic web layer is genuinely low priority compared to landing page performance and geo-targeting precision.

The operators who benefit from understanding this stack now are those building programmatic interfaces, ecommerce flows, or service APIs they want AI agents to discover and use. If your business model involves tools or capabilities that other agents will eventually need to find, ARD is relevant infrastructure to begin understanding. For everyone else, a properly structured performance ads program with clean attribution will outperform any amount of machine-readable metadata until the agentic discovery layer matures significantly.

The One Rule That Hasn’t Changed

The schema.org adoption cycle is a useful template. Early adopters got visible ranking signals. Widespread use and gaming followed. The platforms absorbed the signal and reduced its direct weight. The lesson is not that structured data was worthless — it is that its value is highest at the beginning of adoption, before the signal gets normalized into baseline expectation.

OKF and LLMs.txt are in that early window now. If you can implement them cheaply on appropriate content — developer docs, API references, structured internal knowledge bases — there may be a marginal early-mover signal worth capturing. Do not architect your entire content program around them. Do not convert your site to markdown. Do not stop your paid acquisition to focus on this.

A precision targeting audit of your existing paid channels will almost certainly surface more recoverable margin than any investment in the agentic web stack at current maturity levels. The second layer of the web is real and it is being built. The third head — where agentic systems fully diverge from the human web — is still speculative. Operators should understand the stack, implement what is cheap and appropriate, and keep the main acquisition engine running.

Originally reported by Search Engine Journal, July 2026.