AI Chatbot for Your Business Website: Connect It to Your Knowledge Base Without Hurting the Customer Experience

AI chatbot for a company website: the short answer

An AI chatbot won't harm the customer experience if it answers only from validated internal sources, acknowledges its limits and hands the conversation to a human as soon as it has doubts. Without these three conditions, it becomes a

Which knowledge base sources should you connect first?

Start with the content your customers consult most and that rarely changes: FAQs, delivery and return policies, then product pages. Technical documentation comes next, once the pilot is stable.

Before connecting anything, clean up your knowledge base:

• remove duplicates and contradictory versions;
• withdraw expired offers and old prices;
• assign an owner to each source, responsible for keeping it up to date.

A chatbot repeats what it reads. If two pages give different delivery times, it will hesitate or pick one at random.

It's better to have a few reliable sources than many approximate ones. You can expand later, one source at a time.

How does the chatbot find the right answer?

The chatbot doesn't answer from memory. It relies on retrieval-augmented generation, or RAG.

The principle works in two steps:

1. Retrieval. For each question, the system finds the passages in your documents that are closest to the topic.
2. Writing. The model drafts an answer from those passages alone, not from its general knowledge.

The result: the answer follows your delivery terms, not an average of what's out there on the internet.

This reduces made-up answers, but doesn't eliminate them. If the search returns an off-topic passage, or if the source is outdated, the answer will be wrong. That's why cleaning up beforehand matters.

Safeguards: what does the chatbot do when it doesn't know?

A reliable chatbot knows how to say "I don't know." Four rules make this possible.

• Limited scope: it answers only from validated sources, never from its general knowledge.
• Admitting uncertainty: if no relevant passage is found, it says so rather than improvising.
• Cited source: each answer links to the page used, which the customer can check.
• Escalation to a human: the conversation passes to an agent with the full history, so the customer doesn't have to repeat themselves.

The trade-off is real. A strict confidence threshold reduces errors but increases handoffs. A looser threshold automates more, with more risk. Set it by topic: strict for prices and returns, looser for general information.

Site integration: widget, loading speed and conversion

A well-integrated chatbot doesn't weigh on site performance. It loads after the essential content, stays discreet and triggers at the right moment.

• Deferred loading: the script only starts once the page is interactive, or on the first click on the widget. Product pages display as fast as before.
• Discreet placement: an icon at the bottom of the page, with no window hiding the buy button or the cart.
• Contextual triggering: help offered on the delivery page, with no automatic opening at checkout.
• Accessibility: keyboard navigation, sufficient contrast, easy to close.

Measure the loading time before and after adding it.

Measuring impact: the metrics to track from the first month

Track four metrics from the first week:

• Resolution rate: share of conversations closed without human intervention.
• Handoff rate: share of conversations passed to an agent.
• Satisfaction: a rating requested at the end of the conversation.
• Unanswered questions: topics the chatbot couldn't handle.

Don't look for a universal benchmark. Set your own starting point, then compare week after week.

Each week, read through the unanswered questions. For each recurring topic, write or fix a source page, then check that the metric improves.

An iterative rollout plan in four steps

An iterative rollout limits risk and makes each step measurable.

1. Scoping: define the scope, the validated sources and the rules for handing off to a human.
2. Pilot: launch on a single topic, such as delivery, and analyze the conversations.
3. Expansion: add one source at a time, as long as the metrics stay stable.
4. Maintenance: assign an owner to each piece of content and work through the unanswered questions.

A technical partner handles execution, integration and monitoring, so the chatbot stays reliable over time.