Invisible by Accident

The café is in the answer. The machine shop is not.

That was the first thing that fell out of a small experiment I ran on the island I live on. Sixty-one local businesses – makers, places to eat, guesthouses – put through nine sourcing questions of the kind a customer or a buyer would actually type into an AI assistant. Where do you get a good lunch around here. Who does CNC machining on the island.

Sixty of them were covered by at least one query. Eighteen came back in an answer. Forty-two did not exist.

Broken out, the pattern is sharper than the total. Tourism: seven of fifteen. Places to eat: six of eleven. Manufacturers: five of thirty-four.

The assistant can find you a table. It cannot find you a supplier.

One caveat I owe you before any of this counts. This was a first pass: one assistant, one run, nine queries, scored by the same session that built the index. These engines are not deterministic, and a clean multi-engine rerun is the obvious next step. The gap is too wide for that to reverse, but it is softer than the round numbers make it look.

The detail that turned this from a statistic into a story was the file called llms.txt – a plain text file that tells an AI what a business is and what it does. Thirteen of the thirty-five makers on the island have one.

Every single one of those thirteen is a food, drink or craft producer with a webshop. Their shop platform generated the file automatically, as a feature, at no cost and with nobody asking.

Among the businesses where a machine-readable enquiry is worth real money – machinery and metalworking, stone, infrastructure – the count is zero. Not one.

So the island's machine-readability is real, and it was distributed entirely by accident. The businesses that got it were the ones who happened to buy a certain kind of online shop. The businesses who would profit most from being found by an agent that sources parts have nothing, because nothing came free with what they bought.

Then the number that reframed the whole project for me: of sixty-one businesses, none block AI crawlers. Zero user-agent blocks. Zero robots rules naming any of the AI bots. Nobody opted out.

Forty-two businesses are invisible to the machines and not one of them made a decision about it.

This is, as far as I can tell, the actual shape of local AI adoption everywhere – not a refusal, and not a skills gap.

When the US census asked businesses with fewer than five employees why they had no plans to use AI, the top answer was not cost, not privacy, not lack of skills. Eighty-two percent said AI was not applicable to their business. Lack of knowledge came in at under seven percent. Privacy at six.

The small-business regulator's own reading of that is unusually blunt: either AI companies design for larger businesses, or the owners are too busy running the day to learn about it – and either way, the industry has to work harder to reach them.

There is a quieter structural fact underneath. The European statistical survey that measures business AI adoption only covers firms with ten or more employees. The one-person craft producer, the two-person guesthouse, the village bakery are not undercounted in the official picture of AI adoption. They are absent from it. We are making policy about a long tail we do not measure.

Now the part where I have to argue against my own project.

If you have read that AI is destroying local search, the number behind it is probably wrong. The most-quoted figure says AI Overviews appear on around 68 percent of local searches. That average is dragged upward by informational queries, where they show up more than nine times in ten. On queries with genuine local intent – the ones that actually walk a customer through a door – AI Overviews appeared 15 percent of the time and the old map pack 93 percent.

So nobody on this island is losing their Tuesday lunch trade to a chatbot. Not this year.

The honest case is slower and less dramatic. The damage shows up first in the questions before the decision – what to do on a rainy afternoon here, who can machine this bracket – and those are exactly the questions where a name either appears or does not. Invisibility is not a cost today. It is a compounding one.

Which is where the project stopped being technical.

I built a public page scoring every business. Then I had to rebuild it, because in a place this size you cannot publish a league table of your neighbours. They meet at the harbour. Their kids share a classroom.

So the ranking came out. Five businesses are celebrated by name at the top. Everyone else appears alphabetically with a broad band and no number. Your actual score is visible only on your own card. Then I threw away the colour scheme – dark blue, alarm red – and rebuilt it in warm amber and coral, because red on a café owner's name does not read as diagnostic. It reads as a verdict.

Praise publicly. Diagnose privately. That constraint was not a design preference. It was the price of being allowed to do it at all.

And it goes further than taste. Where I live, cold-emailing a business you have no relationship with is illegal, and so is a cold message on LinkedIn. This cannot spread as a campaign. It can only spread as a letter, then a phone call, then a conversation.

I assumed that was a handicap. The evidence says it might be the mechanism.

The largest study of rural small-business AI adoption I could find – a panel of nearly ten thousand firms – puts rural adoption at 7.0 percent against urban at 9.4. A real gap, and a much smaller one than the noise implies. Its conclusion is that a rural location is not a barrier at all. The variable that actually predicts adoption is network membership. Not broadband. Not capital. Who you are connected to.

A separate study of 4.6 million businesses, using bank transactions rather than surveys, found the same thing from another angle. The dividing line in AI adoption is not revenue. It is whether there is a second person in the business. The smallest firms with an employee adopt at higher rates than the largest firms without one.

Read those together and the picture inverts. Adoption is not a function of money or infrastructure. It happens when there is somebody to talk to about it.

A small island is not short of that. It is made of it.

The tools did not skip this place. Nobody here ever said no. They were never asked.