A web site fabricated from pages fingers a machine half of what it wants. It might probably learn every web page by itself, however nothing on the web site tells it how the pages relate, and nothing tells it whether or not what it’s studying remains to be true. These are two separate gaps. For some time I solely had my eye on the primary one.
Google’s Open Data Format is the closest factor I’ve discovered to a method of each. In June, I constructed an OKF bundle for this web site, eight linked markdown recordsdata, one idea every. I mentioned on the time it was a guess on a reader that doesn’t exist but, as a result of no AI agent reads a web site’s OKF bundle at the moment, and possibly none ever will. That’s nonetheless true. What I didn’t say plainly sufficient is that the reader was by no means the purpose.
The web page itself was by no means a given. We took it from print, the web site as a rack of journal spreads, a homepage, sections, a stack of neat little bundles, and bought so used to it that we stopped seeing it was ever a alternative. When technical web optimization specialist Jono Alderson got here on the podcast, he tied the form all of us copy to Google’s imprint of what a web site ought to be, and known as most of what’s left a brochure, a husk that was once a supply. A machine by no means agreed to any of that. It doesn’t need your homepage; it needs the which means and the way the items join. The web page was all the time half for us, and possibly, all alongside, a bit foolish.
The Worth Proper Now Is The Mirror, Not The Reader
Proper now, OKF’s worth has nothing to do with a machine studying your bundle. It could be value extra later, and if brokers ever do learn revealed bundles, a web site that already has one is forward. However at the moment the worth is a special factor: constructing one forces you to have a look at your web site, and your small business, the best way a machine has to.
That sounds mushy till you strive it. To jot down a bundle, it’s important to say what you truly know, which ideas matter, and the way every connects to the others. Most web sites have by no means declared any of that. They’ve pages, and the relationships between them stay solely in somebody’s head. The second you write them down as a graph a machine can comply with, you discover the components you have been obscure about.
I discovered mine in June. Writing the bundle made me state, in a kind a machine might comply with, that WebMCP sits beneath Machine-First Structure and that llms.txt is identical form of guess as the remainder of the identification work. These relationships have been actual, and so they have been nowhere on the web site. They lived in my head, and a machine studying my pages one by one would by no means have recovered them. Declaring them was not busywork. It was the primary time the form of what No Hacks is aware of existed exterior of me.
So, very first thing straight: Anybody treating OKF as “implement this and get cited” has it backwards. Nothing cites it. It’s a self-audit sporting an export format’s garments. Decide it as an outline of what a machine wants from you, not a product you deploy.
What v0.2 Added Is The Second Hole
The June bundle dealt with the primary hole: construction. Every idea linked to the others, so a machine might see the relations a flat copy of the pages by no means states out loud.
v0.2, launched on July 25, provides the second hole: belief. It places a small set of fields on each idea: the place the content material got here from, who produced it and when, who verified it and whether or not that was a human or a machine, when it ought to be handled as stale, and its lifecycle standing. None of it’s thrilling. It’s the metadata that ought to have been there all alongside, the half that lets a machine ask not solely what you mentioned however whether or not it might probably belief it.
The design alternative value noting is the one Google omitted. It information the indicators and doesn’t compute a belief rating. Its reasoning, in its personal phrases: a rating “is subjective, doesn’t port throughout shoppers, and goes stale the second it’s written.” The patron reads the uncooked indicators and decides for itself. That’s the proper name, and it’s the reverse of the transfer the remainder of the trade retains making, handing you a single quantity it produced about its personal work and asking you to belief it. A recorded sign you possibly can examine beats a rating you’re advised to consider.
The Fields You Can’t Fill In Are The Level
I upgraded my very own bundle to v0.2 to see what it takes. Including the fields is fast. Filling them in actually isn’t, and that’s the place the mirror will get sharp.
The sector that made me cease was the staleness date, the purpose a machine ought to cease trusting an idea with no recent verify. It forces a choice you possibly can in any other case dodge: how briskly does this truly age. My llms.txt idea bought three months, as a result of the story round it adjustments continually. My Machine-First Structure idea bought a yr, as a result of the framework doesn’t transfer a lot. v0.1 let me faux each idea was equally present. v0.2 made me say out loud which of them rot.
The one a part of v0.2 I couldn’t use is value admitting, as a result of it tells you what the format is basically for. v0.2 provides a technique to publish a quantity together with the sanctioned technique to recompute it. My bundle is ideas, not metrics, so there was nothing to attest. That’s high-quality. It’s constructed for information pipelines, not a data map, and pretending in any other case can be the box-ticking the entire train is supposed to reveal.
OKF does nothing at the moment. No agent reads it, form of like llms.txt, no rating is dependent upon it, no customer reveals up due to it. It’s a promising course and, proper now, an empty one. Jono’s darker level from that very same episode nonetheless holds: Most web sites are already husks for a human viewers that’s leaving, and no one, not Google, not OpenAI, has a working reply. OKF reveals you the issue with a readability nothing else does. Fixing it’s a completely different job, and nobody has carried out it.
So, right here is the transfer, and it doesn’t require touching OKF in any respect: Take your most vital web page and ask the three questions v0.2 asks of each idea. The place did this declare come from. When does it cease being true. Who stands behind it, an individual or a course of. A web page that claims “right here is the easiest way to do X” and can’t reply any of the three is a web page a machine has no motive to belief, and actually, neither does a cautious particular person. The fields you can not fill in are the map of what your content material can not defend.
Proper now, the explanation to care about OKF has nothing to do with a machine studying your bundle. The format is the clearest mirror anybody has constructed of what a web site fabricated from pages leaves out. Don’t implement it. Let it present you what you can not but defend, and begin there.
Extra Sources:
This put up was initially revealed on No Hacks.
Featured Picture: Roman Samborskyi/Shutterstock
