Every chatbot demo looks good. These six questions separate the ones that will work on your website from the ones that will embarrass you.

1. What does it do when my pages do not cover the question?

"It says it cannot find that" is the right answer. Anything else - "it uses its general knowledge", "it gives a helpful response" - means it guesses, and it will guess in front of a customer about a refund policy you never wrote.

2. Does it show where the answer came from?

A link a visitor can open is the only practical way to check an answer. It also tells you, when something is wrong, exactly which page to fix. Without it a wrong answer is a mystery you get to hunt for.

3. What exactly does it read, and who decides?

You should see the list of pages, tick and untick them, and read the site again when you change it. If the tool decides for you, you cannot fix a bad answer by fixing a page - which is the only fix that scales.

4. How do I find out when it got something wrong?

Unanswered and poorly answered questions should be counted and shown to you. A chatbot with no such list is a chatbot nobody can improve; you will hear about the failures from customers instead.

5. What happens to my content and my visitors' questions?

Your pages should be used to answer your visitors and nothing else - not to train anybody's model. Ask how long questions are kept, what is stored about a visitor, and whether an IP address is recorded. If your own privacy policy has to describe this, you need the answers in writing.

6. What does it cost when we grow?

Ask what the price does at twice the traffic, not just today. Find out what is counted - pages, conversations, messages, seats - and what happens when you pass the limit: does it stop, queue, or quietly charge?

Then try it on your own site

A demo on somebody else's content proves nothing. Point it at your pages, ask the six questions your customers ask, and then ask three it cannot possibly know. Read the answers as a customer would, not as somebody hoping it works.

Ten minutes of that is worth more than any feature list, and it is the only test that catches the failure that matters: a confident answer to a question your website never answered.

A seventh, if you are cautious

What happens if you leave? Ask whether you can export the questions your visitors asked and the answers given. Those logs are the record of what your customers wanted to know, and they are worth more after a year than they look on day one.

The related question is what removal takes. For most website chatbots it is deleting one script tag, which is a healthy answer - it means the tool never got its hooks into your site. Anything requiring a developer to remove is worth a second look before it goes in.