The short answer
Adding an AI chatbot to a website takes one line of HTML and about ten minutes — you create an account, point the assistant at your existing pages so it has something true to answer from, and paste a script tag before </body>. The work that decides whether it is any good is not the installation: it is choosing what the assistant is allowed to talk about, what it does when it does not know, and where a captured lead ends up.
The mechanics are genuinely trivial now and every vendor’s onboarding is much the same. So this guide spends one section on installation and the rest on the four decisions that actually determine whether the thing helps or embarrasses you.
The installation, briefly
- Give it your content. Point it at your site and let it crawl, or upload documents. This is the assistant’s only source of truth about your business — a model knows nothing about you, so anything it says unprompted about your refund policy is invention.
- Check what it learned. Ask it five questions you know the answers to, including one that is deliberately awkward. This step takes four minutes and catches most disasters.
- Paste the snippet. One script tag before the closing body tag. On WordPress, Shopify, Webflow, Framer or Squarespace this goes in the custom-code or footer-scripts box.
- Watch the first fifty conversations. Not the metrics — the actual transcripts. You will learn more in half an hour than from a month of dashboards.
Decision 1 — what is it for?
“Answer customer questions” is not a scope, it is a hope. Pick one job first:
- Deflection. Answer the questions your support inbox gets forty times a week so people stop emailing them.
- Qualification. Work out whether a visitor is worth a sales conversation, and capture enough to have one.
- Booking. Get someone onto a calendar without the four-email dance.
These want different assistants. A deflection assistant that keeps trying to book a meeting is irritating; a booking assistant that answers detailed policy questions is doing someone else’s job. Start with one, add the second once the first is boring.
Decision 2 — what happens when it does not know?
This is the single most important configuration choice and most people never make it deliberately.
An assistant that guesses is worse than no assistant, because a confident wrong answer about your returns policy creates a support ticket and a disappointed customer, which is the opposite of both things you wanted. An assistant that says “I do not have that — here is how to reach a person” costs you one deflection and buys you trust.
Test this explicitly. Ask your assistant something genuinely outside its knowledge and see what it does. If it improvises, fix that before you launch, not after.
Decision 3 — where does a lead go?
A lead sitting in a chatbot dashboard is a lead nobody is looking at. The question to answer before launch is: which human sees this, in which tool, within how long?
Route it into wherever your team already works — Slack, a spreadsheet, your CRM, WhatsApp — rather than adding a new tab to check. The integrations we support exist for exactly this reason, and the choice matters more than which one you pick.
Decision 4 — how does someone reach a person?
Every assistant needs an exit. Two rules that are easy to get wrong: make the route to a human visible rather than hidden behind three failed attempts, and make handover absolute — once a person takes over, nothing automated should reply into that conversation again. An assistant that interrupts a colleague mid-sentence is worse than one that was never installed.
What it costs
Most vendors, us included, meter conversations rather than seats. Free tiers are common and real. The number worth checking is what happens at the top of the plan — whether you are throttled, charged overage, or cut off — because that is the number you meet on your best week, not your average one. Ours is on the pricing page, including the overage rate.
Two things to get right before launch
- Tell people it is an AI. Since 2 August 2026 the EU AI Act requires it at the first interaction, and it is your vendor’s legal duty rather than yours — but check they have actually done it. California and Utah have comparable rules.
- Update your privacy notice. It has to mention the assistant, what it collects and where conversations go. The full checklist is in running an AI chatbot without breaking privacy law.
How to tell in two weeks whether it worked
Do not start with deflection rate — it is trivially gamed by an assistant that answers everything confidently and wrongly. Start with these:
- Read fifty transcripts. Unglamorous, irreplaceable.
- Count the questions it could not answer, and ask whether the answer should have existed on your site anyway. Usually it should. That list is the most valuable output of the first fortnight.
- Check whether support volume moved on the specific topics you targeted, not overall.
If you want to see the shape of this before committing to anything, the demo runs on our own documentation, and how it works explains the retrieval step in more detail.
