The Web Is Growing A Second Layer – Almost A Third Head

Jul 07, 2026 04:00 PM - 1 week ago 10300

The past fewer weeks person been noisy. Google shipped thing called the Open Knowledge Format. Then Google Developers announced the Agentic Resource Discovery (ARD) specification.

Meanwhile, each SEO LinkedIn provender is lit up pinch personification either declaring markdown the early of the web aliases explaining why you should disregard each of it.

The truth, arsenic per usual, sits location much absorbing than either camp.

The web is processing a parallel machine-readable infrastructure (MCP/WebMCP, OKF, ARD, LLMs.txt…) and SEOs who understand what each furniture really does, alternatively than treating it each arsenic “AI SEO” aliases a metallic bullet, will make amended decisions astir wherever to walk their time.

First: The Layer Cake

There are astatine slightest six chopped things being discussed nether the umbrella of “making your tract AI-ready.” They beryllium astatine different layers and service different purposes:

Crawlable HTML Pages: Still the foundation. Nothing has changed here. Everything other sits connected top.
Schema.org/Structured Data: Semantic hints baked into HTML that show machines explicitly what a page is about. It is, successful essence, a vocabulary.
LLMs.txt: Essentially a navigation file. Its intent is to fundamentally show an AI supplier that’s already connected your tract which pages matter. But arsenic John Mueller puts it connected the Search Off the Record podcast:

“If personification is already connected your website, possibly immoderate benignant of automated strategy is helpful. Where if it goes, I want to spell to Martin’s Splitt and bargain a photograph, past the LLM strategy tin spell to your website and tin look around, like, really do you bargain a photograph? Maybe he has immoderate guidelines for maine arsenic an supplier for buying photographs. That benignant of makes sense.”

MCP/WebMCP: Before ARD came into play, we were presented pinch different solution for the situation of interoperability. An MCP, successful its simplest explanation, is simply a standardized measurement for an AI to link to your services to extract knowledge aliases return action. WebMCP, arsenic the sanction itself suggests, gives websites a measurement to prosecute pinch agents directly. WebMCP is for unrecorded browser interactions connected a webpage; MCP is for devices and services beyond the page.
Open Knowledge Format (OKF): A bundle of markdown files pinch YAML frontmatter.
Agentic Resource Discovery (ARD): A caller unfastened spec for really agents find and verify tools, skills, and different agents crossed the web. Here, the attraction is not your content; it’s your capabilities.

For ecommerce, there’s different furniture worthy naming separately – the product feed – rather perchance the early of unit discovery.

Each furniture does thing different.

I could support adding to this list; there’s a caller furniture popping up each 5 minutes. I’m stopping here. It’s ballooning.

What OKF Actually Is (And Isn’t)

Google published the OKF spec quietly, bolted to a rebrand of Dataplex into Knowledge Catalog.

The format itself is almost disarmingly simple: a directory of markdown files, each pinch a mini YAML header declaring a type, title, description, resource, and immoderate tags. The files nexus to each different for illustration immoderate markdown archive would. That’s it.

As Google’s ain blog puts it, OKF is “just markdown, conscionable files, conscionable YAML frontmatter.”

SEO Suganthan Mohanadasan has a clear breakdown of this. He describes OKF arsenic 1 level successful a stack that now includes sitemap.xml (which URLs exist), LLMs.txt (which pages you astir want read), and OKF (the room itself). They stack alternatively than compete.

The disorder sets successful not erstwhile you look astatine what OKF is, but what it does and successful which furniture of the agentic and hunt mayhem it sits.

In my mind, OKF is not a retrieval system. It doesn’t switch crawling. And, personally, I do not spot a early wherever AI systems nary longer ingest monolithic amounts of HTML aliases wherever hunt and RAG are not a multistep analyzable pipeline that consists of self-reported and “unbased” signals.

Any self-reported strategy tin and will beryllium gamed. So reasoning you tin conscionable slam a bunch of markdown files connected your tract and beryllium THE preferred prime successful retrieval and find is far-fetched.

OKF is simply a higher-signal root among many. It whitethorn trim parsing costs and amended awesome quality, but it doesn’t switch existing pipelines.

It’s besides worthy being honorable here: OKF was built for information teams, not trading sites.

It arrived arsenic a measurement to stock soul knowledge, i.e., array schemas, runbooks, metric definitions, betwixt AI agents wrong organizations. Pointing it astatine a nationalist website to maine seems a spot for illustration we are yet again repurposing.

Francois Vanderseypen makes the astir precise constituent astir what OKF really is and isn’t: a directed chart of markdown files is simply a web of documents, not a knowledge chart (at slightest not successful its purest sense). A existent KG has explicit, queryable, typed relations. OKF leaves what a nexus implies wholly up to the producer, and an LLM still has to infer the semantics each azygous clip it sounds it.

For me, this points to the crux of really I understand the web and what we do arsenic SEOs. OKF doesn’t alteration the stack. It adds 1 much input into it.

It’s not a shortcut. There are nary shortcuts.

The Schema.org Parallel, And Why It Matters

One of the patterns to understand present is the 1 Schema.org already went through.

Structured information followed a predictable arc:

Adoption – ranking boost – wide usage (and gaming) – level learning – reduced dependency arsenic a ranking signal.

FAQ schema had a infinitesimal successful SERPs, past Google discontinued the FAQ rich | result. The platforms study from the signals, fold the lessons into the algorithm, and the definitive markup becomes little necessary.

OKF and LLMs.txt whitethorn travel the aforesaid path. They’re astir valuable early, arsenic clear signals successful a world wherever AI systems are still learning to parse the web.

Over time, if the formats work, the systems learn. Explicit markup becomes redundant aliases remains a verification layer. For example, successful ecommerce, successful particular, schema and provender alignment has go much and much important. Another notch successful the telephone for co-ownership of the merchandise provender betwixt SEO and paid teams!

There’s besides a subtler constituent worthy making present astir the narration betwixt schema.org and discovery. Jarno van Driel’s heavy dive connected merchandise variants successful MCP illustrates this well: For years, Google Search and Google Merchant Center had conflicting system information requirements, forcing publishers to copy markup. Schema.org evolves to adjacent gaps, but it’s slow, it’s complex, and implementation is still often a mess.

Structured information has ne'er been a plug-and-play ranking lever. OKF won’t beryllium either.

Should You Convert Your Site To Markdown?

It’s a large fat nary from me. That doesn’t mean I won’t trial it and use carefully!

And John Mueller said it connected the Search Off the Record podcast:

“When it comes to things for illustration a hunt motor aliases astir apt besides successful generic LLM system, having a website that uses normal HTML for the pages is critical. Because a hunt motor aliases crawler tin conscionable spell to that page. It tin recognise each of the different links that are wrong the website.”

The structural accusation successful HTML – nav links, footers, header hierarchies, soul links – is really crawlers understand your site’s shape. Markdown files portion each of that out. You’d beryllium breaking find successful bid to marginally amended instrumentality readability of individual pages.

Recently, connected LinkedIn, I moreover saw a piece of research showing really “Your navigation mightiness beryllium eating your LLM (it’s ChatGPT Deep Research successful fact) reference budget.” Interesting findings, but please don’t region your navigation to “save immoderate tokens”!

Jono Alderson makes this constituent brilliantly: “A page is not conscionable a instrumentality for words. It’s an editorial artifact.” Hierarchy, emphasis, placement, what comes first, what’s prominent, what’s tucked successful a footnote … these aren’t beautiful decorations for humans. “They are signals astir meaning.”

“When you flatten a page into markdown, you don’t conscionable region clutter. You region judgment, and you region context.” And the infinitesimal you people a machine-only representation, you’ve created a 2nd campaigner type of reality.

The boring hole still works: Semantic HTML, clear structure, sensible hierarchy, contented that exists erstwhile the page loads.

John Mueller covers the markdown statement extensively successful the podcast: The parallel versions problem, the move rendering lessons we already learned the difficult way, and why maintaining a protector type of your tract for AI doubles your attraction load and creates a debugging nightmare cipher will show you about.

The 1 objection Mueller carves retired is developer documentation:

“If you person thing for illustration developer documentation, where, again, if the supplier aliases the LLM strategy already knows astir your website and the personification says, really do I usethis API? Then if you springiness the LLM strategy a Markdown file, it’s a batch easier for it to understand.”

Now, I tin decidedly spot a straightforward usage lawsuit there.

What ARD Is Actually Doing

The Agentic Resource Discovery specification, announced by Google connected June 17, 2026, is simply a different beast entirely. It arrived only a mates of days down OKF, not a coincidence, and is already making immense waves.

The problem ARD solves is simply a coordination one. Right now, an supplier has to beryllium wired to each tool, MCP server, aliases API it uses earlier it tin do thing pinch it.

That useful erstwhile you’re connecting a fistful of known services. It stops scaling the infinitesimal the number of disposable capabilities grows beyond what immoderate squad tin pre-configure by hand.

ARD moves that find retired of setup and into runtime. The supplier finds what it needs erstwhile it needs it, alternatively than only knowing what it was told astir successful advance.

It’s built connected 2 primitives:

Catalogs: An ai-catalog.json record hosted connected your domain, describing your disposable capabilities (MCP servers, A2A agents, OpenAPI tools). Ownership of the domain acts arsenic the cryptographic instauration for personality and trust.
Registries: Search engines for the agentic web. They crawl catalogs, scale them, and return matching capabilities pinch the metadata needed to verify the patient earlier connecting.

If OKF is astir packaging knowledge for consumption, ARD is astir advertizing capabilities for connection.

These are parallel efforts astatine different layers of the emerging agentic stack. Both shipped wrong inches of each different and now adopted pinch the velocity of ray by immoderate very large players successful the game, i.e., Hugging Face and their Discover Tool.

It’s perchance a much pragmatic stake than the general logic furniture that came earlier it and ne'er reached web scale. Time will tell.

A Gap Worth Watching

Within days of some specs shipping, a contributor opened companion issues connected the ARD and OKF repos pointing retired thing basal was missing: There’s no agreed media type for an OKF bundle, truthful a catalog tin database 1 but can’t really admit it arsenic OKF without sniffing the contents.

In the meantime, publishers are already advertizing bundles successful accumulation utilizing their ain interim types, which, arsenic the rumor itself notes, won’t work together pinch each other.

On the look of it, this looks for illustration a mini ask, conscionable a petition for a shared label.

After a spot of a dive into this peculiar rabbit hole, it turns retired that’s rather normal practice. Waiting for afloat statement earlier anyone ships thing is precisely really a spec dies successful committee, and shipping accelerated and patching arsenic existent take surfaces is an age-old strategy.

Application/json itself wasn’t formally registered until 2006, astir 5 years aft JSON was already successful wide, informal use. Nobody worried astir that, because the costs of the explanation being unsettled was low: A parser mightiness cull thing aliases autumn backmost ungracefully.

But OKF is different, because what happens aft the fetch is different. The artifact down the explanation is simply a bundle an autonomous supplier is meant to ingest, verify, and perchance enactment on, wrong a find strategy built specifically for agent-to-agent and agent-to-tool connection. Get the type incorrect here, aliases time off an supplier to infer it, and the consequence isn’t a parse error; it’s a strategy acting connected thing it shouldn’t person trusted, pinch nary 1 checking the consequence first.

I wonderment astir the consequence progressive successful settling this later alternatively than sooner successful this case. I conjecture it depends connected really accelerated it gets resolved comparative to really accelerated take runs up of it.

What This Means If You’re An SEO

A fewer honorable conclusions and my existent thinking:

For astir trading and contented sites, not overmuch has changed. HTML, well-structured for humans, is still the correct foundation. A contact-us shape and a cleanable tract architecture will service you amended than immoderate OKF bundle ever will. Discovery still depends connected links, authority, personification signals … and indexing.

LLMs.txt is simply a signpost, not an SEO tool. It’s useful for helping an supplier navigate wrong your tract erstwhile it’s already there. It very apt doesn’t make a large quality successful really agents find you successful the first place. And, astir apt ne'er will.

MCP/WebMCP. Neither is urgent for astir trading sites today, but if you’re building thing pinch programmatic interfaces aliases ecommerce flows you want agents to navigate, this is the guidance the infrastructure is heading.

OKF makes a batch of consciousness if you’re sitting connected system soul knowledge, i.e., documentation, API references, merchandise specs … and you want to make it easier for agents to consume. The free OKF generator Suganthan built will nutrient a bundle and springiness you a chart position of your soul nexus building arsenic a broadside benefit. The structural audit unsocial seems worthy it. But I will beryllium moving it connected my website first, not connected my client’s website.

ARD is worthy watching if you’re building services pinch programmatic interfaces. If you person tools, agents, aliases APIs you want discoverable by different agents, ARD is the emerging modular for really that gets done. Just know the personality layer underneath it, what an supplier is really looking astatine erstwhile it finds your catalog entry, is still being settled successful existent time, truthful I’d dainty this arsenic infrastructure to watch intimately alternatively than build captious paths connected conscionable yet.

The schema take rhythm mightiness repeat. These formats are astir valuable now, arsenic early signals. Implement them if you tin do it cheaply. Don’t build your strategy astir them holding worth everlastingly and don’t slope connected them arsenic a metallic bullet.

Ultimately, beryllium alert of the shiny things – if your institution has bigger food to fry, i.e., a unspeakable website, a marque nary 1 knows aliases cares for, an assemblage you don’t understand … past woody pinch this first earlier you get caught up successful immoderate of these caller shiny things.

The Underlying Shift

What each of this points to is simply a web that’s genuinely increasing a 2nd furniture aliases a 3rd head, 1 written for machines alongside the 1 written for browsers and humans.

Sitemap.xml told crawlers which URLs existed. Robots.txt told them wherever not to go. LLMs.txt, OKF, and ARD are akin infrastructure for agentic systems: navigation hints, contented packaging, and capacity discovery.

None of it is mandatory today. None of it replaces coagulated HTML, charismatic content, sensible structure, aliases the point that really sits underneath each of it: a marque worthy finding.

But the SEOs who understand what each furniture really does, alternatively than treating it arsenic a azygous undifferentiated “AI SEO” category, will make amended bets connected wherever to walk their time.

My money is connected the 2nd layer, a parallel infrastructure written for machines, not a replacement for what already exists.

The 3rd caput scenario, wherever agentic systems afloat diverge from the quality web, would require a different group of bets than immoderate of america are presently making.

Big acknowledgment to Jarno van Driel, Jono Alderson, Chris Green, Suganthan Mohanadasan, Kristine Schachinger, Gianluca Fiorelli, Victor Pan, Renee Bigelow (and anyone other I’ve missed) for immoderate superb discussions connected this taxable complete the past fewer weeks.

More Resources: