How it works
The mechanics, without the hand-waving.
What the assistant reads, when it decides to hand a visitor over, exactly what lands on your phone, and what you can turn off.
01 — What it knows
It reads your site, then you fill the gaps.
Give it your website address and it crawls the site — following links three levels deep, up to fifty pages — and keeps the text. That gets you a useful assistant on day one without writing anything.
Then add the part your site never says out loud: pricing rules, what you don't take on, the lead time you quote, the answer you give forty times a week. Plain text in a box — no FAQ builder, no categories to maintain.
Knowledge base
Persona
02 — When it hands over
Four ways a conversation becomes your problem.
You decide which of these are live. When one fires, the assistant asks for a name and a number in the flow of the conversation — no form drops out of the sky — and then texts you.
Words that mean money
You give it the words that matter in your trade — "quote", "emergency", "same day", a product you make good margin on. The moment a visitor types one, the conversation is a lead and you hear about it, however early in the exchange it happened.
Someone still talking after N turns
A visitor who is six questions deep is not browsing. Set the number of exchanges after which the assistant should stop trying to close it alone and get you involved. It is the trigger that catches the interested people who never say a magic word.
then hand over
The assistant decides
The model is told what a lead looks like for your business and can raise its hand on its own — for the visitor describing a job, comparing you against someone else, or asking a question that only a person should answer. This is the one that catches what your keyword list didn't.
"We're replacing two doors on a rental, need it done before the 14th."
Every conversation, no exceptions
For businesses where anyone who types at all is worth a call. The assistant still answers the question properly first — it just always collects contact details and always texts you. Blunt, and the right setting more often than you'd think.
Whichever fires, you get one text per conversation — never a second alert for the same visitor.
03 — What you get
A summary, not a transcript.
Nobody reads a twelve-message transcript on a phone at a job site. The assistant writes what the visitor actually wanted, in a couple of lines, with the name and number above it — enough to decide whether to call back now or after lunch.
Alerts can go to one phone or several — in rotation so nobody carries every lead, or all at once when speed matters more than fairness.
Asked whether you install in Ogden. Confirmed Wasatch Front coverage, roughly a week out. Wants a quote call — mentioned two doors on a rental, needs it before the 14th.
04 — What you can switch off
Settings that admit the AI isn't always wanted.
Every one of these is per widget, not per account — so the assistant can run your after-hours storefront while your support page stays a plain form.
Let the assistant answer around the clock, only when you're closed, or not at all — in which case the widget captures details and nothing else.
Pop open after a delay, or wait to be clicked. Suppress it on mobile, or show it only on your home page — useful when the widget competes with a checkout.
Conversations are capped at 20 exchanges, abusive traffic is rate-limited before it reaches the model, and if the AI ever fails the widget falls back to a contact form rather than breaking.
Point it at your site and see what it says.
The crawl takes a few minutes. After that you can ask it your own customers' questions and judge the answers yourself.
Start your 14-day trial