Guide

AI chatbot for website: how to add one the right way

Adding an AI chatbot for website visitors? Pick one job, ground answers in your own content, set clear limits and plan the handoff to your team.

Key takeaways
  • Launch the assistant on one measurable job, such as your top five pre-sale questions, before you add booking or lead capture.
  • Rewrite the pages behind your most common questions so each price, hour and policy is stated once, in full sentences.
  • Before launch, test with 30 questions including ones your site doesn't answer, and confirm every handoff reaches a named person within a promised time.

Most people who set out to add an AI chatbot for website visitors start in the wrong place. They compare vendors, paste a snippet into the footer and then spend the next month wondering why the thing feels either useless or faintly risky. The installation takes an afternoon. What decides whether the chatbot helps or embarrasses you is everything around the installation: which job it’s hired to do, what it’s allowed to say, and what happens at the edge of what it knows.

This guide walks through that work in the order it’s worth doing it. You don’t need a developer to start, and none of it is glamorous. It’s the difference between a chat widget that answers real questions and one that confidently invents answers in your name.

Pick one job for your AI chatbot for website

Every vendor page lists the same menu: answers, bookings, lead capture, translation, summaries, follow-ups. It’s tempting to switch on all of it. Don’t. A chatbot that tries six jobs on day one gets judged on six things and usually does none of them well, because each job exposes a different gap in your content and your processes.

Pick one job and make it measurable. Picture a physiotherapy clinic with a two-person front desk. In an inbox like that, it’s common for a large share of messages to be the same three questions: whether the clinic takes a particular health plan, what a first appointment costs, and where to park. A chatbot hired to answer those questions can be live within a week, and you can judge it by a number you already track, which is how many of those messages stop arriving by email.

Booking, lead capture and after-hours coverage come later, once the basic answers are trustworthy. Each one adds a new failure mode, and you want to understand the first before you stack the second on top of it.

Audit the questions you already get

Before you touch any software, export everything your visitors have asked you in the last 90 days. That means the contact form, the general inbox, any live chat logs, the notes your receptionist scribbles from phone calls, and direct messages from social accounts. Three months is usually enough to show the seasonal pattern without drowning you in old data.

Read through the messages and sort them into clusters. Most businesses end up with fifteen to twenty-five. Then count each cluster and rank it. Ranking by volume alone is a mistake, though. A question that comes right before a purchase, such as “can you do this on a Saturday?”, deserves more attention than a question about your holiday closures, even if the holiday question arrives more often.

The most useful column in the spreadsheet is the one that asks whether your website already answers the question clearly. Here is the shape of a finished audit. The numbers are illustrative, not data from a real business:

Question cluster Typical wording Share of messages (example) Clear answer on site today?
Pricing “How much is a first visit?” 22% Partly: price is in a PDF
Availability “Do you have anything this week?” 18% No
Location and parking “Where exactly are you?” 11% Yes
Insurance and payment “Do you take my plan?” 14% No
Hours and holidays “Are you open on Monday?” 9% Yes, but out of date
Cancellation policy “What if I can’t make it?” 6% Buried on a sub-page

The rows marked “no” or “partly” are your content work list. They also tell you what the chatbot can’t answer yet, which is the honest starting point for any launch.

Make your content answerable

A grounded assistant doesn’t carry a memory of your business. When a visitor asks something, it searches your site for the passages most relevant to the question and writes its answer from those passages. The quality of the answer is therefore capped by the quality of the passage. If the price lives in an image, or in a PDF titled “Fees 2024 final v3”, or is spread across three pages with three slightly different phrasings, the search has a hard time and the answer comes out vague or wrong.

The fix is to write the way a good receptionist would speak. Each fact should appear once, in a complete sentence, with its subject named. Here is an invented example of one paragraph rewritten:

Before: “Our rates are competitive and we’re happy to discuss what works for you. See the details below.”

After: “A first physiotherapy session costs £65 and lasts 60 minutes. Follow-up sessions are £45 for 30 minutes. If you have a private health plan, we can tell you whether we’re a provider before your first visit.”

The second version does two things the first can’t. It gives the assistant an answer it can repeat, and it gives a human reader the same answer without scrolling. Where you genuinely can’t give an exact price, state the range and the condition that moves it. “Assessments start at £65; complex cases can take longer and are quoted after the first session” is far more useful to a visitor, and far safer for the assistant, than silence.

Pay attention to things that change: opening hours over bank holidays, a seasonal price, a closed location. Put a “last reviewed” date in the content owner’s calendar. Stale content is the most common reason a well-built assistant gives a wrong answer, and usually the assistant is faithfully repeating a page somebody forgot about.

Write for the question, not the page: give each common question its own short heading and answer it in the first sentence underneath. The section that opens with the answer is the one an assistant, and a skimming visitor, can use fastest.

Set the boundaries before launch

An assistant needs three lists before it goes live. The first is what it’s for. The second is what it must never do. The third is when it must stop and bring in a person.

For the first list, be concrete. “Answer questions about our services” is too loose. “Answer questions about services, prices, hours, location and first-visit preparation; book first appointments into the calendar; capture a name and contact for everything else” is something you can test.

The second list is where most businesses are too casual. A reasonable starting list for a service business looks like this:

  • No promises about outcomes. It can describe a service, but it doesn’t say a treatment will work, a repair will hold or a project will finish on a given date.
  • No discounts or exceptions unless you have published them. Negotiation belongs with a person.
  • No advice in regulated areas such as medical, legal or financial guidance beyond what your own pages already say.
  • No commitments you haven’t made about availability, delivery or stock. It should check the calendar or say it doesn’t know.

The third list is the handoff, which gets its own section below.

Anything it can read, it can repeat: before you connect the assistant, remove draft pages, last year's promotions and anything marked "TBC". A grounded assistant is only as honest as the pages behind it, and it has no way to know that a banner was meant to expire.

What the assistant should say when it doesn’t know

This is the part people underestimate. A visitor will ask something you never considered, and the assistant needs a good default. The right default is short and honest. It says it doesn’t have that detail, offers to pass the question to someone, and asks for a contact method. “I don’t have that detail, but I can pass your question to the team so someone replies today” is a good response. A confident guess is not.

Plan the handoff as carefully as the answers

A handoff is the moment the chatbot stops answering and a person takes over. Most failed deployments fail here rather than on the answers. The chat promises that someone will get back to the visitor, and then nobody is watching the inbox.

Design the handoff around four questions:

  1. What triggers it? Typical triggers are an explicit request for a person, a complaint, a cancellation, a question outside the content, or a high-value request such as a large quote.
  2. What does the person receive? At minimum: the visitor’s name and a way to reach them, a one-sentence summary of what they wanted, the page they were on, and whether they flagged it as urgent. A team member should never have to ask the visitor to repeat the whole story.
  3. Who receives it, and how? Route by type. A booking question goes to the front desk. A billing dispute goes to whoever handles billing. A shared inbox that three people half-watch is the most common weak link.
  4. What does the visitor hear? Give a time you can actually keep. “Someone will reply by 9am on the next working day” is honest. “Someone will be in touch soon” is a promise nobody can audit.

Treat the handoff as a service level. For the first month, measure how many handoffs were answered within the promised time. If the number is poor, change the routing or the promise, not the wording of the chat.

Test it like a skeptical customer

Before launch, write thirty test questions and try them yourself. Divide them into four groups of roughly equal size, because each group catches a different failure.

  • Easy questions that the content answers clearly. The assistant should answer these correctly every time, and in substance the same way across attempts.
  • Edge questions that the content answers only partly, such as “Do you do weekend appointments for children under eight?” The right behaviour is an honest partial answer plus a handoff.
  • Off-topic questions, such as asking for a recipe or a competitor’s price. It should decline briefly and steer back to what it’s for.
  • Pressure questions designed to extract a promise: “Can you guarantee I’ll be seen today?” or “Give me 20% off if I book now.” These are where you find out whether the boundaries you set are real.

Ask the question you can't answer: the most useful test is one your site genuinely doesn't cover. If the assistant admits the gap and routes it, your design is working. If it invents a plausible reply, fix the boundary before anything else goes live.

Run each test at least twice, in a second phrasing. Assistants are fluent, so the same question asked differently can produce a noticeably different level of confidence. Keep a log of every failure with its fix. Typical fixes are a rewritten content passage, a new boundary rule, or a changed handoff trigger.

The failure modes to expect are fairly predictable. Holiday hours get given from the regular timetable. A discount is offered because a promotion page was still live. Availability is described from a calendar that was never connected. A question about a competitor gets a long answer because nothing told the assistant to stop.

Launch, then tune weekly for the first month

Launch with a schedule, not a switch. Put a fifteen-minute review on the calendar every week for the first four weeks. Each review has four parts.

First, read the list of unanswered questions. Each one is either a content fix or a boundary decision. Most weeks the list will suggest two or three page edits, and those edits are where the real quality gains come from.

Second, check the handoff log for volume, response times and reasons. A cluster of handoffs for a question you thought was answered means the answer isn’t as clear as you believed.

Third, sample ten random conversations and read them in full. Numbers hide tone. A transcript where the answer was correct but curt, or accurate but sent to the wrong person, is worth more than a dashboard tile.

Fourth, count the outcomes you care about: bookings made through the chat, enquiries captured, and questions resolved without a person. Be careful with the last one. A question “resolved” by a visitor who simply gave up is not a win.

Expect the first two weeks to produce most of your content fixes. After that the volume drops, and the remaining work is mostly about new services, new staff and seasonal changes.

1Pick one jobAnswer pre-sale questions, book, or capture leads.2Audit the questionsCluster 90 days of real messages, rank by volume.3Ground answers in your contentState each fact once, clearly, in full sentences.4Test, hand off, tuneProbe with hard questions, route people, review weekly.
Four steps, in this order: pick one job, audit real questions, ground answers in accurate content, then test, hand off and tune.

What to take away

An AI chatbot for website visitors is a small technical task wrapped in a larger editorial one. Its usefulness is set by three decisions you make before it talks to anyone: the single job, the boundaries, and the handoff. The answers follow from the content, and the content is something you already own and can keep improving.

If you want to see how this plays out, the Voxy live demo lets you put the questions in this article to the assistant yourself, including the ones it shouldn’t answer. Our guide to training an assistant on your content goes deeper on the content side, and the grounding article explains why a good assistant should say “I don’t know” in the first place.

Voxy follows the same order of work. It answers from your own pages, books into your calendar when the job calls for it, and hands off to your team when a person is needed. If you’d like to try it on your own content, you can start a free 14-day trial with no credit card.