The standard pitch goes like this. Point the tool at your website, it reads every page, and now visitors can ask questions instead of hunting. It takes an afternoon to set up and costs less than a phone line.
All of that is true and most of these end up sitting unused in the corner of the screen. The reason is simple enough that it rarely gets said out loud. A visitor who wants to know your hours can find your hours. A visitor who wants to know whether you can come out Saturday cannot, and that is the question the bot usually cannot answer either, because the website does not know.
So the bot answers the easy things, which nobody needed help with, and deflects the hard things, which is the whole reason the person opened it.
The test any chatbot has to pass
One question. Can it tell a visitor something they could not have found by scrolling?
That is not a rhetorical standard. It is the practical line between a bot that gets used and a bot that gets closed. Everything genuinely useful in this category sits on the far side of it: whether there is an opening Thursday, whether you cover that zip code, what a job like theirs usually runs, whether the thing they own is something you service, whether their insurance is one you work with.
Each of those requires the bot to reach something beyond the page. A calendar. A service area. A price table. A list of what you do and do not take. None of it is exotic, and almost none of it is in the default setup.
Reading your site is the floor, not the product
Training a bot on your own content is worth doing. It is just not a feature, it is a prerequisite, and treating it as the finished product is why so many of these underperform.
There is a second problem with content-only bots that takes a few weeks to surface. They inherit every gap in the site. If your pages never say plainly whether you take a given insurance, the bot will not say it either, and unlike a page it will be asked directly and repeatedly. What you get is a log of the questions your website fails to answer, which is genuinely valuable, but only if somebody reads it.
We treat that log as a content brief. The questions people type into a bot are the questions that belong in the pages themselves, written plainly, where search engines and answer engines can also see them. A bot that quietly improves your site over a year has paid for itself twice.
Three jobs a chatbot is actually good at
Qualifying before anyone's time is spent
Most service businesses take inquiries that were never going to be jobs. Outside the service area, wrong kind of work, budget that does not reach the floor price. A short set of questions asked up front, politely, filters a meaningful share of that before it reaches a human, and it does it without the awkwardness of a person asking.
Booking, when it can see a real calendar
This is where the connection matters most. A bot that offers times it has confirmed are open closes the loop in one sitting. A bot that collects a request someone else schedules later has moved the work rather than removed it, and introduced a delay where the customer expected an answer.
The question you answer twenty times a week
Every business has one. Do you take this plan. Do you come out this far. Is the first visit chargeable. How long does it take. What do I need to have ready. Those are worth automating not because they are hard but because they are constant, and every one handled is a few minutes back for somebody.
What it should refuse to do
A chatbot that will answer anything is a liability with a friendly icon. The model underneath is built to produce a plausible response, which means an unconstrained bot will invent a price, a policy or a clinical reassurance rather than admit it does not know.
For medical practices this is not a theoretical risk. A bot should not be interpreting symptoms, should not be telling anyone what their plan covers, and should not be touching anything specific to an individual patient. The line we draw is that it can describe what a practice offers and help somebody get to a person. It does not practise, and it does not adjudicate coverage.
For everyone else the same principle applies with lower stakes. Decide what it may commit you to. If it quotes a number, that number is going to be held against you by somebody who screenshotted it, and arguing that the bot was mistaken is not a position you want to be in.
- Anything a customer could reasonably treat as a binding quote
- Anything clinical, diagnostic or specific to one person's situation
- Anything about legal exposure, eligibility or what a policy covers
- Anything it is inferring rather than reading from a source you control
Chat is not always the right shape
It is worth asking whether a chat window is the right container at all. On mobile, where most of this traffic is, a bubble that covers a third of the screen is often just in the way, and plenty of visitors would rather tap a button than compose a sentence.
Sometimes the better answer is a short guided flow with real options. Sometimes it is a phone number that gets answered, or a text back on every missed call, because the customer wanted to talk and chat was a detour. Sometimes it is simply writing the answer on the page where it belongs.
We have talked businesses out of a chatbot more than once. When the real problem is that nobody picks up the phone, adding a window to the website does not fix it, it just gives the inquiry a second place to die.
The overlap with how AI answers now work
There is a connection between this and something happening off your site entirely. When someone asks an assistant which businesses do a particular thing in their city, the answer is assembled from pages that state things plainly enough to be quoted.
The content that makes a chatbot useful is the same content that makes a page quotable. Clear, specific, self-contained answers to real questions. Doing this work once serves both, which is why we tend to treat a chatbot project and AI search optimization as the same body of work rather than two line items.
What we would build
Start from the three questions your team answers most and the one thing a visitor cannot find on the site. Connect the bot to whatever holds the real answer. Draw a clear line around what it may say. Read the transcripts every month and move the good answers onto actual pages.
That is a smaller project than most vendors describe and it holds up better. We build these into the sites we run for businesses and medical practices in South Florida, as part of the website itself rather than a widget bolted on afterwards. If you have one now and are not sure it is doing anything, we will look at the transcripts with you and say so either way.