The things an agent cannot guess.

A wrong recommendation made with confidence is worse than a wrong guess

A regular who walks into a quiet countryside pub on a Tuesday afternoon knows, without being told, that muddy boots and a dog are fine, that the back room fills up for the Sunday quiz, and that the kitchen stops taking orders early on weeknights. None of it is written on the website. A first-time visitor picks most of it up anyway, from the chalkboard, the regulars, the general shape of the room. An AI agent recommending that pub to someone planning a walk has none of that. It only knows what's been declared somewhere it can read, and "somewhere it can read" is a much smaller set of facts than "a person could work out just by looking around."

Say the same pub isn't really set up for a large group booking, tables pushed together in a room built for couples and pairs of walkers. The landlord knows this instantly and steers big groups towards the function room down the road. An AI has no equivalent instinct. Asked to recommend somewhere for a birthday party of fourteen, it has nothing to stop it suggesting the pub with total confidence, unless that limitation has been declared as plainly as the opening hours.

A person who guesses wrong in the room can correct course. They ask a question, notice the atmosphere, feel the fit isn't right and leave before it costs them anything. An AI making a recommendation in advance doesn't get that chance, and neither does the customer acting on it. That's the part worth sitting with: the confidence itself is the risk. A pub that's genuinely not right for a large group needs to say so as clearly as what it is right for, because an AI repeating a guess with total confidence does more damage than a person guessing out loud ever could.

Most businesses only write down what they're proud of

Nobody puts "not really suited to large groups" on their own homepage, even when it's exactly the fact an AI recommending on their behalf most needs. A short, plain, sometimes slightly unflattering fact, stated once, does more work for an AI than a whole paragraph of atmosphere. "Not licensed for children after 8pm" tells an assistant something useful in six words. Three paragraphs about the warm, welcoming atmosphere tell it nothing it can act on, and writing more of that kind of copy was never going to close the gap.

What counts as an answer

The test worth applying to any piece of local knowledge is simple: would a new member of staff, on their first shift, need to be told this, or would they just pick it up? If it's the kind of thing that takes a new starter a few weeks to absorb, that's exactly the kind of fact an AI has no way to absorb either, however many times it reads the website. It has to be told directly, the same way a new starter eventually gets told, except it needs telling once, clearly, rather than picked up slowly over a few shifts, one confused customer at a time.

Deciding what a person would eventually just know, and saying it plainly instead of leaving it to be picked up, is a shorter job than most businesses expect. That's the specific gap Selfe exists to close: turning the things a regular knows, or a landlord tells you without being asked, into facts an AI can check directly, rather than leaving them to be absorbed by whoever happens to visit enough times to notice.

Isn't this just writing more content?

No, usually the opposite. The facts that matter most here are short and specific, not more paragraphs of description.

What if being honest about what we're not right for puts people off?

It puts off the people who'd have been disappointed anyway. Declaring a weak fit prevents the booking that ends in a complaint, not the one that would have gone well.