I let an AI build an SEO site by itself. Here's exactly where it hit the wall.

Let me confess what NomadRoutes actually was, because the shelf label is a lie of omission. On the surface it's a tidy little site: real monthly budgets, internet speeds, and a "Nomad Score" for 118 remote-work cities, refreshed every month. Useful enough. But I've spent a good while in digital marketing, and I didn't build it because I was itching to compare rent in Kathmandu and Chiang Mai.

I built it to answer a question that's been nagging at everyone in my field: can an AI, more or less on its own, produce a website that search engines actually pick up — the kind of thing that ranks in Google, or gets surfaced by whatever we're calling it this month, AIO, GEO, generative engine optimization, pick your acronym. I wanted to watch a system try to beat a system. So I gave the AI the wheel: the topic, the architecture, the structure — I barely steered.

I wasn't really interested in digital nomads. I was interested in machine versus machine — and in staying out of its way to see what happened.

For context, "digital nomad" is just this era's word for freelancing. Remote work made cross-border contracting genuinely viable, and there's a real, growing allergy to being chained to a "proper job." Fertile territory, lots of search demand. A good test bed. So — how far did the machine get?

01What the AI actually got right

More than I expected, honestly. It picked a real topic with real demand. It stood up a coherent site. And it had one instinct I thought was genuinely sound: instead of just listing raw numbers, it tried to synthesize its own metric — a single Nomad Score blending the various factors that might affect a nomad's life in a given city. That's the right idea. A number that didn't exist before. If you're going to compete for attention, having your own proprietary index is exactly the move.

The problem was never the idea. The problem was the depth.

02Where it broke: a score with no soul

Look closely at that Nomad Score and it falls apart. It's built from rent, an average food cost, internet speed — and that's about it. But people aren't spreadsheet rows. Humans are complicated, and the score didn't model a single one of those complications.

Am I a man or a woman? That already changes things. Am I married or single — and who decided married people can't be nomads? If I've got a family, I don't want a one-bed; I need a family suite. And if I've got a kitchen, I'm probably buying groceries and cooking, so the "average food cost" is wrong for me too. What if I want to rent a car? What if part of the point is to actually see the place?

and honestly — if you just want a change of scenery at your desk, apply for a normal job. that's not nomadism.

It computed something. It just computed something almost meaningless. A number that precise and that shallow is arguably worse than no number, because it looks like it means something. The fix I can picture is beautiful and enormous: simulate all these personas, run each through every city, build a real database underneath. That would be worth ranking for. The AI stopped three orders of magnitude short of it.

03Where it broke: it didn't have time to think

Here's the part I keep turning over. Watch how many sites it surveys in one research pass. Watch how quickly it identifies a button when it's driving a browser. My gut says the same thing every time: it needs far more time than we give it.

My theory — and it's only a hunch — is that to avoid making you wait, and to live within a usage budget, the AI simply can't gather enough. It surveys a thin slice, moves on, and produces something plausible on top of an under-researched foundation. It had ideas. It just didn't have depth, because depth costs time, and time costs tokens, and nobody was paying for enough of either.

04The lesson I didn't expect

Before anything else, spend a huge amount of time — and tokens — gathering and verifying the data. Maybe make that its own separate project entirely.

That's the whole game, I think. The build is the easy, visible part. The moat is the research nobody sees. First, you have to feed the AI properly — more care, more iteration, more tokens — or it stays at "fine." Second, and this stung: you probably need real domain expertise yourself to make it good. Wikipedia is famously spartan yet ranks like a monster — so "polished" isn't the answer. Knowing what actually matters is. You can't outsource the judgment.

05Why I stopped (the unglamorous truth)

I'd love to give you a dramatic ending. I don't have one. I waited a few days, opened Google Search Console, and saw a flat line at zero. And I decided I didn't want to keep burning usage on a machine-versus-machine fight I was barely in. Honestly? I just ran out of budget to keep feeding it. Ha.

Not a heroic pivot. A guy looked at a zero, shrugged, and went to do something else. Which is, I suspect, how most projects actually end — not with failure, but with a quiet reallocation of attention.

NomadRoutes still sits there, correct and current and completely unvisited. There's a poetry to it: a tool built to help people get found couldn't find its own way to the top. But now you know what it really was — not a nomad site that failed, but an experiment that told me exactly where the wall is. And walls, once you know where they are, are just the next thing to get over.