How to build a chatbot for your website: a step-by-step method

Start from your customers’ real questions, not the ones you imagine. Re-read the emails you receive, the messages from your contact form and the notes taken on the phone: you will see the same topics come up again and again (opening hours, prices, lead times, service area, procedures). These are what the chatbot should handle first.

Next, define what a successful conversation should produce: a sufficient answer, contact details collected, an appointment booked, a pointer to the right page. Without an expected outcome, there is no way of knowing whether the chatbot works.

Finally, write down what it must not do: give medical or legal advice, promise a firm price on a quote that requires an assessment, handle a sensitive complaint. This negative scope is just as important as the rest, because it determines when the chatbot must stop.

A chatbot only knows what you give it. Gather your website pages, your FAQ and your documents (terms, published price lists, service descriptions), and write instructions for anything that is not written down anywhere. Check that these sources do not contradict each other and that they are up to date: out-of-date information becomes a wrong answer.

Set the tone: formal or informal address, short sentences, level of detail. Give the chatbot a name and a welcome message that make it clear it is an automated assistant and what it can do. Because a chatbot often offers free-text input, users may share sensitive data with it (health, political opinions, religious affiliation) without you having planned for it: in that case the CNIL, France’s data protection authority, asks for a warning to be displayed, before the chatbot is used at all, advising users not to share such data, and for this data to be purged immediately or at least regularly.

Finally, take care over the widget itself: it should open and close easily, be readable on mobile and remain usable with a keyboard. The Web Content Accessibility Guidelines (WCAG 2.2), published by the W3C, serve as the reference for checking these points.

Decide in which situations the conversation should go to a person: an explicit request from the visitor, an out-of-scope question, an answer that cannot be found in the content, a sensitive topic, dissatisfaction. In all these cases, the chatbot collects contact details and a summary of the request, so nobody has to ask everything again.

Plan who receives the alert, by what means and within what timeframe, including evenings and weekends. Give visitors a timeframe you can actually meet: promising an immediate callback that nobody provides damages trust more than an honest timeframe.

For the decisions that matter, stay in control. In France, the CNIL points out that a conversation with a chatbot without human intervention cannot, on its own, lead to a decision with significant consequences for a person, such as refusing credit.

Without code, off-the-shelf platforms let you import your content, write your instructions and publish a widget by pasting an embed code on your website. This is the fastest route for a small or medium-sized business, provided you check what the tool does with questions it cannot answer and how it passes requests on.

With code, you assemble a language model, a search across your documents, an interface and connections to your tools yourself. You gain control, but you take on hosting, security, maintenance and testing every time the model or your content changes.

Either way, sort out the data question before going live. According to the CNIL, a cookie set only when the user activates the chatbot, and used solely to make it work, does not require consent; if it is set before activation or used for anything else, consent is required. Also set a retention period for conversations that suits their purpose, and find out where they are hosted.

Before publishing, test with a list of real questions, then with difficult ones: clumsy wording, typos, off-topic questions, contradictory requests, a visitor who wants to talk to someone. Have it tested by someone who does not know the tool: they will ask the questions you did not anticipate.

Once it is live, track a few indicators, but the right ones: the share of conversations that reach the expected outcome (answer, contact details, appointment), unanswered questions, and requests passed on to the team. The unanswered questions are your list of content to add.

Regularly review a sample of conversations. It is the only way to spot an answer that is correct in substance but misunderstood, or an instruction the chatbot applies too broadly.

The same mistakes keep coming up: too broad a scope from the outset, instead of a few topics handled well; content never updated after launch; a chatbot that hides the fact that it is automated; a window that opens by itself and covers the page; a form that asks for more information than necessary; a callback promise the team cannot keep.

The most costly is having no way out: a visitor going round in circles, unable to leave their contact details or reach a person, may leave the website, and their request is then lost.

If you would rather not build everything yourself, Alyas offers a chat agent whose name, colour and welcome message can be customised. It relies on your website, your documents and your instructions; if it does not know, it says so and offers to pass the request on. Before publishing the widget, you message it from your dashboard, as a customer would; it is then installed with a single line of code pasted on your website, restricted to your domains. Visitors’ contact details and needs are saved to your leads, an appointment can be booked if Google Calendar or Cal.com is connected, and WhatsApp, Messenger or Instagram can be added once connected from your dashboard, subject to your accounts’ permissions and availability validation. When a request needs to go to your team, the conversation is flagged and you are alerted by email.