How to build an AI chatbot to automate website sales

A website chatbot is the chat window in the corner of a site. A good one answers a visitor's question from your own pages, then does something with the answer: books a call, collects contact details, or points to the right page. In Zapier you build one in three stages: train it, set it up to act, then style it and put it live.

The video above is the whole build, start to finish, followed by five real chatbots tested from the customer's side. This is the written version, with the prompts you can copy.

What can a website chatbot do besides answer questions?#

Answering questions is the floor. On an online store, a chatbot can act like a sales associate: the customer asks for a product, the bot replies in plain language with a few options, and the customer adds one to the cart without leaving the chat.

For a service business, the useful jobs are different:

• Qualify. Ask a few questions to see whether a visitor is a fit, and only offer a call if they are.
• Book. Let the visitor pick a time inside the chat, then send the calendar invite and reminders.
• Log support issues. Sort an incoming problem and file it in your ticketing system.
• Hand off. When it cannot answer or cannot act, follow a rule you set and bring you in.

Do most small businesses already have one?#

Fewer than you would think. Before filming, we checked the homepages of 2,227 businesses listed in the member directories of six California chambers of commerce. 2,032 of them loaded. 50 showed a chat window of any kind, and only 9 of those were a bot. Two of the nine belonged to national companies. Most of the rest were "text us" buttons or a message box that goes to the owner's inbox.

What that count misses: we looked at homepages only, a widget that appears after a scroll or a delay would not show up, and some sites block automated visits. Even allowing for that, a small business with a working chatbot is still the exception.

Which tool should you use?#

There are many chatbot builders. We used Zapier Chatbots for this build, partly on cost and partly because Zapier's whole product is connecting one tool to another, which matters once the chatbot starts taking action. Check Zapier's pricing page for current plans, since they change.

The business we built for is IntelliChoice, a professional services firm run by Ana. Her site has a lot of pages, and every visitor is being steered toward one thing: a 15-minute fit call where she learns what they need and whether it makes sense to work together. That is a good shape for a chatbot. Visitors have questions, the answers are spread across the site, and there is one clear next step.

To start: create a Zapier account, go to Automations, then Chatbots, and click Create. You get two choices, build from scratch or create with AI. We built from scratch so every setting is visible.

The screen has three parts. The tabs down the left control the chatbot. The middle shows the settings for whichever tab you are on. The right side is a live preview that stays put while you move between tabs, so you can test every change as you make it.

Stage one: train it#

Training decides what the chatbot says and how it says it. It happens on three tabs: Instructions, Knowledge and AI model.

Instructions: rewrite the directive for your business#

The Instructions tab has two fields, the greeting and the directive. The default greeting is fine to keep. The directive is the part to change. It tells the chatbot who it is talking to, what it is for, and what style and tone to use, and the default is written for nobody in particular.

The default is badly matched to your business, but its format is good. So the fastest route is to have an AI assistant (Gemini, Claude or ChatGPT all work) read your website and write a new directive, then pour it into Zapier's format. Two prompts:

I'm creating an AI chatbot for my business website, [your website address]. Read the site and write a directive for the chatbot: who our visitors are, what the chatbot is for, the one next step we want visitors to take, the services and prices it should know about, and the style and tone it should use. List any facts from the site a chatbot would need to answer customer questions well.

The first answer comes back long. That is fine, because it is full of facts the chatbot needs. The second prompt keeps those facts and changes only the shape:

Without changing the language of what you just wrote, reformat it so it matches the structure of this prompt: [paste Zapier's default directive here]

Paste the result into the directive field and read every line before you save. The assistant is working from your website, so anything stale on the site, or anything it inferred, ends up in what your chatbot tells customers.

Knowledge: give it your whole site, and keep it current#

Knowledge is the information the chatbot draws on to answer. Choose web page, paste your address, and switch on the option to crawl the page for sub pages. Your home page alone leaves out the pages that hold most of the answers: services, resources, contact. IntelliChoice's site had 17 pages, and we selected all of them.

Then set a schedule for the crawl to run again, either for one page or for all of them. When you change a price or add a service, the chatbot picks it up on the next run instead of quoting the old one.

AI model#

This tab has two settings: which model the chatbot runs on, and how it responds with that model. We left the response setting on low and it worked well, but test it on your own questions before you settle.

Test it#

In the preview, we asked: "Do you offer any bookkeeping services?" It answered from the site, included the pricing, and offered a link to book a call. That is the bar for a trained chatbot: the answer, plus a next step.

Stage two: set it up to act#

This stage is what lets the chatbot do something with an answer instead of only giving one. It happens on the Logic and Integrations tabs.

Collect leads#

Collect leads asks the visitor for their contact details inside the chat. You choose the fields, where the answers are stored (a Zapier table), and when it asks: at the start of the conversation, after a few messages, or when the visitor uses certain words. You also choose the format: an embedded form (email, phone, company), or questions asked one at a time as part of the conversation.

We first set it to ask at the start, then changed it to appear only when a visitor says something like human, touch, get in touch, contact or email. Asking a person who just wants an answer for their phone number first is the fastest way to lose them, which the examples further down show.

Schedule a meeting#

Schedule a meeting holds your booking link. We pasted in Ana's Calendly link, and from then on a visitor who asks about a call can book one inside the chat instead of being sent to another page.

Suggestions and link buttons#

Suggestions are ready-made questions that appear at the start of a conversation or after each reply, for visitors who open the chat and do not know what to type. Zapier preloads a few. With them switched on, "Can I schedule a call?" appeared, and clicking it brought up the booking calendar.

Link button sends the visitor a button to a page, on your site or anywhere else, when they use a word you choose. Set it to fire on "services" or the names of your services, point it at your services page, and pick an icon.

Automations: what happens after the conversation#

The last logic option runs a Zap, Zapier's name for an automation, when a conversation ends. The trigger is "chatbot conversation completed", and after it you add any steps you like:

• Email yourself the transcript through Gmail.
• Have ChatGPT summarize the transcript first, then email the summary.
• Send the conversation to your CRM, HubSpot for example.
• Log every conversation as a row in a Google Sheet.

A Zap button works the same way, except the visitor starts it by clicking a button in the chat. A "Call me" button could, for instance, hand off to an AI voice tool that phones the visitor.

Integrations: the same brain on other channels#

The Integrations tab lets other channels use the same instructions, knowledge and model. If someone messages your business on Facebook Messenger, the chatbot can draft the reply there. Once that is set up, the chatbot is less a widget on your site and more the one place your answers come from, wherever the customer writes in.

Stage three: make it look like yours, and put it live#

Theme#

The theme tab sets the chatbot's colors and look. IntelliChoice's site runs on orange and dark blue, so we copied the colors from the site into a document and entered them on the theme tab, so the chatbot looks like part of the site rather than something bolted on.

Settings: call it an AI assistant#

In settings, the public name is what visitors see. Name it something like "IntelliChoice AI Assistant", and the reason is legal as well as polite. Several states have laws about bots that a customer could mistake for a person:

• Maine requires a clear notice that the customer is not talking to a human whenever a chatbot used in trade could mislead a reasonable consumer into thinking it is one (10 M.R.S. §1500-DD, in force since September 2025).
• Utah requires a business to say it is AI when a customer clearly asks, and requires licensed professionals to say so up front in high-risk interactions, such as financial, legal or medical advice (SB 226, 2025).
• California makes it unlawful to use a bot that hides what it is in order to push a sale, and says disclosing that it is a bot is enough to stay clear of the law (Business and Professions Code §17941).

The wording differs from state to state, and your customers come from more than one. "AI Assistant" in the name covers the common ground in two words. This is not legal advice; if your business gives regulated advice, ask your lawyer what your state requires.

Embed it#

To put it on your site: click Share in the top right corner, open Embed, and copy the code. That code goes into your website's code, the same way you would add any other widget. If someone else manages your site, the copied code is what you send them.

On the IntelliChoice site, the chatbot opened with a question, answered "What does your company do?", and brought up Ana's calendar when we typed "schedule". Then we pushed it with a harder question: "I am the owner of a small grocery store in Los Angeles. Can you help me with my financial strategy?" It wrote an answer tailored to a grocery store, drawn from the instructions and the site, and it knew Ana works in QuickBooks.

What do the good chatbots do differently?#

We tested five chatbots on live business websites, as a customer would. The pattern was clearer than expected. The useful ones give the customer something before they ask the customer for anything, and when they do ask, the question is one they need answered to help.

The California DMV ("Ask DMV", whose assistant is named Miles). The preloaded options are good: they are clearly the questions people bring to the DMV most. But it opens with a large disclaimer, pops out into a separate window instead of sitting inside the page, and when we asked how much it costs to renew a driver's license, it pointed us to the right page without stating the fee. It is a navigation aid, not an answer.

Storelocal Storage. We asked for the largest unit we could rent. It asked for our location first, and that question earned its place: it needed the location to check inventory. Then it named the largest unit, gave the price, offered to show availability, and offered a video of the unit. A useful chatbot.

Quirk Accident & Injury Attorneys. The chat opens on its own and asks what kind of injury you have, then asks you to describe what happened, then collects your name, email, phone, city and date of the incident. It gives no legal opinion at all, and that is the right call. For law firms, tax advisors, therapy practices and funeral homes, where the information is sensitive and a wrong answer can do harm, a chatbot that works as a careful intake form and gets a person on the case is a good use of one.

An assisted living community. We asked when we could visit a grandfather who lives there. It asked us to reword the question. We reworded it, in more detail. It could not answer, offered to put us in touch with customer service, and when we asked how to inquire about a new resident, pointed us to the contact page. The visitor typed three messages and ended up on the contact page, which was one click away before they opened the chat.

A fine art gallery. The widget is not well-designed. It invited us to start chatting, so we asked about buying a painting, then for the opening hours. It asked for our name before answering anything, then a phone number, then an email, and then said a team member would be in touch. It never answered either question. This is an intake form dressed as a conversation, and it is worse than a form, because the visitor was invited to ask something and then ignored. If this is the only job a chatbot will do on your site, a phone number and a plain contact form serve the customer better.

Put the five side by side and the test for your own chatbot comes out of them: when a visitor asks a question, do they get an answer or a next step before you ask them for anything? If yes, the chatbot is helping you. If the first thing it does is ask for a phone number, it is in the way.

Written November 2026. Zapier's screens, plans and labels change; the steps above match them as of this month.