The AI-native rabbit hole I probably didn't need

Most of the rabbit holes I fall into start with a small flash of jealousy at how fast something else can move.

This one started when I watched Claude Design put together a presentation in what felt like a couple of minutes. Not a rough draft, a real presentation, properly laid out, with the kind of visual confidence that would have taken me a full afternoon. And I sat there afterwards, mildly impressed and mildly unsettled, with the obvious thought: if this is what the model can do when given clean structured input, what would it do with a B2B website behind it?

And then the slightly worse thought: should the work site be more “AI native” than it currently is?

I spent half an evening on that question. More than I should have. Less than I might have if I hadn’t caught myself.

The version of the idea that pulled me in goes something like this. Sites are increasingly read by models, not just humans. If the goal of a content-led B2B site is partly to be useful to whatever AI ends up summarising or recommending it, then maybe the structure should be more model-friendly. Plain text knowledge base. Markdown-style chunks. Less ornamental layout, more semantic clarity. Closer to how an Astro markdown blog is structured than how a CMS-driven marketing site is structured.

There’s something to that argument. I’m not going to pretend it’s nonsense. AI-driven retrieval is a real factor now and it’s going to keep being one.

But here’s where I had to pull myself out of it. The site I’m building is already structured around clean content relationships, a typed schema, and clusters that express how topics connect. That structure is genuinely good for both human readers and machine readers. It’s not hostile to AI consumption. It just isn’t optimised for AI consumption at the expense of everything else.

And the “everything else” includes a few things that matter quite a lot. Buyers landing on the site through a sales conversation. People comparing options on a phone. Engineers looking for specific technical details inside a longer page. None of that audience cares whether the markup is unusually friendly to a model. They care whether they can find what they need quickly and whether the experience feels solid.

The argument for AI-readability tends to assume the model is the most important reader. For most B2B sites, it’s not. It’s a reader, increasingly an important one, but it’s still downstream of the people who actually buy things.

The other thing I had to remind myself of is that the existing structure already had a reason. The CMS-driven approach with content modelling and clustering wasn’t picked because it looked nice. It was picked because it solves the actual problem of keeping a content-led site coherent as it grows. Trading that for something flatter and more “AI native” would have been swapping a real solution for a speculative one.

So I closed the tab and went back to the polish work. Half an evening gone. Not entirely wasted, because I think I’m clearer now about why the existing approach is the right one. But also not productive. The kind of detour that comes from watching one tool do one thing impressively and immediately wondering if your whole approach is wrong.

That instinct usually deserves a careful look and then a quiet rejection. This one was no different.