Before you add a chatbot to your website

Adding an AI chat widget to a website is now a few minutes of work. That is exactly why so many of them make a business look worse rather than better.

A bot is only as good as what it was given

These systems answer from the material they can see: your pages, your documents, whatever you connected. When the answer is not in there, a well-built bot says so. A badly configured one produces something plausible instead — in your brand voice, on your website, to your customer.

The failure mode is not silence. It is confident invention, and it is worse than no bot because the customer has no way to tell the difference.

The prerequisite nobody mentions

Before you install anything, write down the twenty questions customers actually ask. Not the ones you wish they asked — the ones your staff answer on the phone every week. Hours. Parking. Deposits. Lead times. What is included. What happens if they cancel. Whether you take the thing they have.

Then check whether your website answers all twenty. It will usually answer about eight.

Fixing that gap is worth more than any chatbot, and it is also the thing that makes a chatbot work, because now there is something correct for it to draw on. We have had clients do this exercise and decide they no longer needed the bot. That is a good outcome, not a failed project.

Set the boundary explicitly

Decide what the bot is not allowed to do, and configure it rather than hoping.

It should not quote prices that vary. It should not promise availability. It should not give advice with legal, financial or health consequences. It should not agree to anything on your behalf.

It should hand over to a person — with the conversation attached, so the customer does not repeat themselves — the moment a question crosses that line. A bot that says "let me get someone who can answer that properly" is doing its job. One that guesses is generating a complaint.

Then watch it for a month

Read the transcripts. All of them, at first.

You will learn two things, and both are worth the exercise on their own. What customers really ask, in their own words, which is rarely what you assumed. And where your site is silent, which is a list of pages to write.

You will also catch the bot being wrong, which is the point of reading them. Every wrong answer is either a missing page or a boundary that needs tightening.

The handover is the part people get wrong

Most chatbot complaints are not about wrong answers. They are about being trapped.

A customer who wants a person and cannot reach one will not conclude that your bot is imperfect. They will conclude that you are avoiding them, and they are half right, because that is usually why the bot was installed.

So make the escape route obvious and available from the first message, not after three failed attempts. Pass the whole conversation across, so nobody has to repeat themselves — repeating yourself to a human after failing with a robot is the specific experience people describe when they say they hate chatbots.

And be honest in the opening line about what it is. "I can answer questions about hours, menus and bookings — for anything else I will get someone" sets an expectation the bot can actually meet. Pretending to be a person does not survive the second question, and the customer remembers being fooled more than being helped.

What it costs when it goes wrong

A bot that invents an answer does not fail neutrally. It fails in your voice, on your website, and the customer has no way to know it was the machine.

The expensive version is not a wrong opening time. It is a quoted price you did not agree to, a promise about a lead time you cannot meet, or an assurance about something with a legal or health dimension. Depending on how it was worded, you may find yourself arguing that your own website did not mean what it said.

This is why the boundary matters more than the accuracy. A bot that refuses ten reasonable questions costs you a little goodwill. One that answers a single question it should have refused can cost considerably more, and you will not know it happened unless somebody is reading the transcripts.

What good looks like on a small site

Concretely, because "add a chatbot" covers two very different things.

The version that works is narrow. It answers from your own pages — hours, parking, what is included, how long a repair takes, whether you cover a particular area. It says plainly when it does not know, and hands over to a real person with the conversation attached so the customer does not repeat themselves. It never quotes a price or promises a date. And it is honest that it is a bot, because customers work that out in two messages anyway and resent having been played with.

The version that does not work tries to be a salesperson. It improvises around gaps, guesses at pricing, and produces confident answers to questions nobody ever wrote down.

The difference is not the technology — both are the same underlying tool. It is entirely in how tightly it was bounded, which is a decision you make, not something you buy.

When not to bother

If you get a handful of enquiries a week and answer them yourself within the hour, a chatbot solves nothing. It adds a layer between you and a customer who was already getting a good experience, and the personal reply was probably part of why they chose you.

The case for a bot is volume, repetition and out-of-hours. If you do not have all three, spend the effort on the twenty questions instead. It is cheaper, it helps every visitor rather than the ones who open the widget, and search engines can read it.