Make the page clearer
Put hours, service areas and booking links where people expect them.
Before adding a chatbot, find the question your website leaves unanswered.
If visitors cannot find your prices, service area or booking link, fix those things first. A chatbot becomes interesting when people need help navigating a genuine choice — and you have reliable information it can use. Here is how to tell the difference.
Choose the lightest useful interaction
Start with what the visitor is trying to do.
Offer a useful form
Ask for the details needed for a quote, booking or next conversation.
Consider a chatbot
Use maintained sources and a clear route to a person when needed.
Look at real questions before choosing a chat widget
Read recent enquiries and note what people were trying to find out. Group the questions by the decision they support: can you help me, what does it cost, how do I book, or where is my order?
A repeated question can reveal missing content. It does not automatically justify a conversational interface. Put opening hours on the page; do not make visitors interview a robot to discover them.
| Visitor need | Often a useful starting point | When chat might help |
|---|---|---|
| Find your hours or service area | Clear page content | When the answer depends on a more specific location or service. |
| Request a straightforward quote | A short, relevant enquiry form | When a conversation helps clarify which details are needed. |
| Choose between several services | Comparison page or guided selector | When people describe their situation in varied language. |
| Check a booking or order | An authenticated lookup or existing portal | Only with an appropriate connection and access checks. |
A chatbot needs a source of truth and a boundary
Before writing its personality, list what it may answer and where those answers come from. Useful sources include current service pages, approved FAQs and maintained policies. Give someone responsibility for keeping them current.
A model can generate a confident answer that is wrong. NIST describes this as confabulation. For a business chatbot, the practical response is to test answers against your sources and make uncertainty useful: offer the relevant page or a clear human handoff.
Keep informational answers separate from actions. Explaining a service is different from changing a booking or promising a price. The latter requires an appropriate system connection and an agreed permission boundary.
What a useful conversation actually looks like
Illustrative example: an independent bicycle workshop is considering a chatbot. These are proposed behaviours, not a deployed client system.
| Visitor asks | A useful response | What it must not assume |
|---|---|---|
| Do you repair electric bikes? | Answer from the maintained service list and explain any listed exclusions. | That every motor or battery brand is supported. |
| Can you fix mine tomorrow? | Explain how to request a slot, or use verified booking availability. | That a free-looking time means a confirmed repair. |
| How much will this noise cost to fix? | Explain the assessment process and link to any published fees. | A diagnosis or exact quote from a vague description. |
| I want to speak to someone | Offer the actual contact route immediately. | That another automated answer is an adequate handoff. |
Make the handoff carry the conversation
If the person needs help, ask only for what the team needs to respond. Where the integration supports it, pass the question and relevant conversation context to the enquiry record, rather than asking the visitor to repeat everything.
Tell the visitor what happens next and use a response window the team can meet. Do not show “connecting you now” unless somebody can actually take over.
A small set of approved FAQ answers can be automated without reviewing each one. For an AI agent taking a consequential action, selected tool calls can instead be routed for approval; n8n documents that capability.
On a phone, the close button matters as much as the answer
Test the widget as a visitor, with one thumb and a real question. It should help people finish their task, while leaving the website usable.
- The launcher leaves the main call to action and cookie controls accessible.
- Opening the keyboard does not hide the send button or the latest answer.
- The close control stays visible, and closing returns the visitor to a sensible place.
- Long answers are readable without sideways scrolling; relevant links are easy to tap.
- Keyboard and screen-reader users can reach the controls and leave the conversation.
- Contact details remain available when the chat service is slow or unavailable.
Judge it by successful next steps
Create a small test set from genuine questions, with personal details removed. Include an ambiguous request, an out-of-scope question and a request for a human. Write down what a helpful answer should contain before running the test.
Then review correct answers, unsupported answers, useful handoffs and unfinished conversations. A long conversation is not automatically a good one. If visitors ask three times for the booking link, the widget has created work.
If most questions are answered better by improving a page, make that improvement. If the chatbot consistently helps people navigate a real choice, you have a concrete reason to develop it further.
Sources
What are visitors struggling to find?
Bring a few common questions and your website URL. We can help decide whether the useful first step is clearer content, a guided form or a chatbot.
Discuss the visitor experienceUpdated