Adding AI Chat to Your Website: A Practical Setup Guide

Most businesses that add AI chat to a website do it in the wrong order. They pick a tool, paste the script, tick the default settings and switch it on the same afternoon. Two weeks later the transcripts are full of the assistant apologising for not knowing things, and the widget gets quietly removed.

The install is the easy part and takes about ten minutes. What decides whether the thing earns anything is the content you put behind it and the rules about when it appears. This is a step-by-step guide to both, and it is deliberately tool-agnostic, because the platform you choose matters far less than the preparation.

Decide what job the chat is doing before you choose a tool#

An assistant that tries to do everything answers everything vaguely. Pick one primary job and design around it.

For most service businesses the job is capturing the enquiry that would otherwise have left. Someone lands on a service page at 8pm, has one question standing between them and contacting you, and there is nobody at the desk. The assistant answers that question and gets a name and a number.

For a business with a heavy support load, the job is different. There the assistant is deflecting repeat questions from existing customers so the phone rings less, and the measure of success is fewer calls rather than more leads.

These two jobs pull the configuration in opposite directions. The first should push toward booking early. The second should answer thoroughly and only escalate when it genuinely cannot help. Choosing one gives you a rule for every configuration decision that follows.

Feed it the twenty questions you already answer every week#

This is the step that separates a useful assistant from an embarrassing one, and it is the step everybody rushes.

Do not point the tool at your website and let it index everything. Website copy is written to persuade, not to answer, so an assistant trained only on it produces confident marketing language when someone asked a specific question about a price. The result reads like a brochure that learned to talk.

Instead, write the content specifically. Sit down and list the questions you actually get on the phone. For a plumbing business that list looks like the areas you cover, the call-out fee, whether you charge for quotes, how fast you can get to an urgent job, what you do not do. Write a direct answer to each in the words you would use on the phone.

Then add the operational facts the assistant needs to avoid making things up.

• The exact suburbs or regions you service, including the ones you deliberately do not.
• Opening hours, and what happens outside them.
• Prices or ranges you are comfortable stating publicly, and an explicit instruction to hand over rather than guess when a price is not listed.
• The situations where it must stop and pass the conversation to a person.

There is a fuller checklist of what to feed an assistant, including the material that makes answers worse, in how to train an AI agent on your business.

Write the opening message like a person, not a greeter#

The default opening message on most platforms is some variant of "Hi there! How can I help you today?" It gets ignored because it asks the visitor to do the work of figuring out what the assistant is capable of.

A better opening does two things at once. It signals what the assistant can actually help with, and it makes the first reply easy. On a service page, something like "Want a rough price or to check we cover your area? Ask away" outperforms a generic greeting, because it tells the visitor the assistant knows about prices and coverage.

Match the opening to the page. Someone on your pricing page has a different question from someone on your about page, and using the same greeting on both wastes the one line you get.

Trigger timing and placement decide whether people hate it#

The fastest way to make visitors resent a chat widget is to open it before they have read anything.

Sensible defaults look like this. Wait until the visitor has been on the page for around twenty to thirty seconds, or has scrolled far enough to show genuine interest, before the invitation expands. Open immediately only when the visitor clicks the bubble themselves. Never re-open automatically after someone has closed it, and never play a sound.

Mobile deserves separate thought, because most of your visitors are on a phone and the widget occupies a much larger share of a small screen. Keep the collapsed state small, position it so it does not sit on top of your call button, and check that it does not cover the price table or the form. Test this on an actual phone rather than a browser window resized to look narrow.

The other rule is that the human option must be visible at all times, not buried behind three failed attempts. Hiding it is the behaviour that generates the worst reviews, and it is covered alongside the rest of the greatest hits in why customers hate your chatbot.

Install it, then break it on purpose#

The technical install is a script tag before the closing body tag, or a plugin if you are on WordPress or Shopify. On most platforms it is genuinely one line. If your site is slow already, check what the widget adds to load time, because a chat script is another third-party request and speed is the thing that quietly caps every other number on your site. There is more on that in what a slow website actually costs.

Before you let it talk to real customers, spend half an hour trying to break it. Ask about a service you do not offer. Ask for a discount. Ask something rude. Ask a question in a way that mixes two topics. Ask for a price on the most unusual job you have ever quoted.

What you are looking for is the assistant inventing an answer. Every confident wrong answer you find in testing is one you do not have to apologise for later. Where it invents, add an explicit rule that it must hand over on that topic instead.

Run it in shadow mode for a few days if the platform supports it, where the assistant drafts responses that you review rather than sending them live. It is the cheapest insurance available.

Set up the measurement before you switch it on#

Most businesses cannot tell you whether their chat widget works, because nothing was tracked when it launched and there is no baseline to compare against.

Configure these events in your analytics on day one.

• Chat opened, split by page, so you learn where the questions concentrate.
• Conversation started, meaning the visitor actually typed something.
• Contact details captured.
• Handed to a human.
• Booking made, if the assistant books.

The number that matters is conversations that produced a contact detail or a booking, divided by conversations started. Track it weekly. If it is low, read twenty transcripts and you will usually find one repeated question the assistant handles badly.

Also watch what happens to your overall site conversion rate, not just chat performance. A widget that produces plenty of chats while form submissions fall by the same amount has moved enquiries around rather than created them. If your baseline numbers are not recorded yet, fix that first using the diagnostic in why your website gets traffic and no enquiries.

Where to go once it is running#

Give it a month before making judgements, then read the transcripts properly. They are the most honest customer research you will ever get, because people ask an assistant things they would feel silly ringing about.

Two upgrades usually follow. The first is qualification, where the assistant asks about timeline and scope so the leads arriving are sorted before they reach you, covered in qualifying leads on your website. The second is connecting chat to the same pipeline as your calls and DMs, so an enquiry gets the same follow-up regardless of where it came in. That wider picture, and where chat sits inside it, is in what makes a website smart.

If you only do one thing this week, write the twenty questions list. Whether or not you ever add AI chat to your website, that document makes your service pages better, and it is the input everything else depends on.