How does the AI learn the facts about your house?

From you, in three ways: it reads your listing and house manual into a structured fact sheet, it asks about the gaps it could not fill, and it learns the rest from your corrections, one question at a time. It never answers from the open internet or from a model's general impression of vacation rentals — if a fact is not on your sheet, the AI does not have it. We run this on a live co-hosted listing, so this answer comes with the step the marketing version leaves out: read the fact sheet once, end to end, before the first guest message. Here is why.

It starts by reading what you already wrote

You have already written most of what guests ask about — in the listing itself, the house manual, the welcome doc you send after booking. So setup starts with a paste. An extractor turns that text into a structured fact sheet: check-in and check-out times, the wifi network and password, parking, amenities, house rules, the quirks like the one lamp with the tricky switch, your local recommendations. When we loaded a live listing, the property record was built straight from the listing's public page — rules, amenities, check-in at 4, checkout at 10, pets allowed. There is no forty-field onboarding form; the form is your own writing.

It tells you what it could not find

Alongside the facts it found, the extractor returns a list of what it could not: no wifi password in the text, no parking notes, no door code — door codes should not be in a public listing anyway. That list drives the setup questions you are asked, so you fill in exactly what is missing instead of everything. A tool that pretended the sheet was complete would not need the list — it would just fail at 2am instead, which is the worse interview.

We watched it misread a wifi fact

During our own smoke test, the input "Wifi network is SmokeTestNet and the password is..." came out with the network name recorded as "SmokeTestNet and the password is" — the parse split in the wrong place. And because every reply is grounded in the sheet, the next wifi question got a draft that repeated the mangled name faithfully. Note what the failure was not: the AI did not invent anything. It answered exactly from its facts; the fact itself was wrong. That is the honest limit of extraction — it is a draft of the truth, very good and not perfect.

Read the sheet once before the first message

So the practice we now treat as non-negotiable on any listing we co-host: before the first real guest message, read every fact on the sheet the way a guest would hear it — wifi, door, parking, check-in, pets. It takes about two minutes, and a mistake caught there costs nothing. The system is layered so a miss still does not reach a guest unreviewed — new accounts start in draft mode, where every reply waits in your inbox for a one-word approval — but the cheap place to catch a bad fact is the sheet, not the hundredth draft.

Why it never reads the open internet

A general model knows plenty about vacation rentals in general and nothing true about yours. Your gate code hours, your pet fee, which parking spot is actually yours — these are unknowable from the outside, and a plausible guess about any of them is worse than silence. So the reply engine is instructed to answer only from your facts and the guest's reservation, never to invent door codes, prices, policies, or amenities, and to route the question to you when the sheet does not answer it. The wifi story above is the price of that rule — a wrong fact in means a wrong fact faithfully out — and the read-it-once habit is the whole cost of paying it.

Then it learns in the order guests teach it

Every held draft arrives with one sentence on what was missing from the sheet. Add the fact once — the pool heater's season, the gate code hours — and that question never stumps it again. Over a season the sheet converges on your actual house, gap by gap, in the order guests hit them, which is a better curriculum than any form you would fill out on day one.

You can watch the first step on your own listing right now: paste it into the demo at /try and look at the fact sheet it builds before you ask a single question. What it found, what it flags as missing, and the one fact you would correct — that is the whole setup, in miniature.