Until a few years ago, most website chatbots were decision trees of pre-written options that often led users to a dead end. Large language models changed that: today, an assistant can understand a visitor's question in natural language and answer based on your content. But a badly set-up bot can also confidently give wrong information about your prices or policies. The difference is in the setup.
Old bots versus a new generation of assistants
Rule-based bots only recognise predefined questions; phrase something slightly differently and you get “Sorry, I didn't understand.” AI-powered assistants interpret a question by its meaning. The reliable ones usually work with an approach called retrieval-augmented generation (RAG): the assistant first searches your documents, pages and FAQs for relevant information, then builds its answer only from that information. This significantly reduces the risk of the model inventing answers from its general knowledge.
When does a chatbot help?
- Your team answers the same questions again and again: delivery times, opening hours, service scope, return terms and so on.
- You get enquiries outside working hours: a question that arrives at night can be answered right away, or captured to pass to the right person, instead of waiting until morning.
- You want to pre-qualify leads: the assistant can ask about needs, budget range and timeline, and hand your sales team a ready summary.
- You have a large catalogue or lots of documentation: it helps visitors find the right product or information.
- You have visitors in several languages: modern models can answer fluently in many languages.
When doesn't it help?
- You get few enquiries: for a handful of questions a month, a well-written FAQ page is cheaper and more reliable.
- Your work involves highly personal or sensitive advice: in areas such as legal, medical or financial decisions, an assistant should at most be a signposting tool.
- Your content isn't ready: an assistant is only as good as the information you give it. Incomplete or contradictory content means incomplete or contradictory answers.
- Nobody owns it: without someone reviewing conversations and updating content, the bot gradually starts giving outdated information.
Photo: Ice Family / Unsplash
Six principles for doing it right
- Ground it in your own content. The assistant should build answers only from approved content and, where possible, link to the page the information came from. That builds trust and takes visitors deeper into your site.
- Draw clear boundaries. The assistant should politely decline out-of-scope questions and never invent a price, date or commitment that isn't in your content. A bot that can say “I don't know, let me connect you with our team” is worth far more than one that tries to answer everything.
- Make handing over to a human easy. At any point in the conversation there should be a way to reach a person: a contact form, e-mail or live support during working hours. If the assistant passes on a summary of the conversation, the customer doesn't have to start over.
- Design for privacy from the start. Tell visitors clearly that they're talking to an AI assistant; the EU AI Act requires this kind of transparency. Know which provider conversation data goes to, how long it's kept and whether it's used for model training, and explain it in your privacy notice. See our guide to a GDPR- and KVKK-compliant website for the details.
- Don't let it slow your site down. The chat widget should load when the user interacts with it or after the page has loaded, not during the initial load. A heavy chat script can drag down your Core Web Vitals. On mobile, it should sit as a small button that doesn't cover the content.
- Measure and improve. Review conversations regularly: which questions couldn't the assistant answer, which answers caused frustration, how many conversations became enquiries? Unanswered questions are the clearest map of the gaps in your content.
Chatbots and SEO
Chatbot answers aren't indexed by search engines, so a bot is no substitute for good content. But it's a valuable source of data: the questions visitors ask the bot are the real list of questions your site needs to answer. Turning them into FAQ sections, service pages and blog posts improves both your organic visibility and your chances of being cited in AI-powered search. We cover this in getting cited by AI search.
Frequently asked questions
Will a chatbot replace customer service staff?
No, and it shouldn't. A well-built assistant takes on repetitive questions so your team can spend its time on conversations that genuinely need a human touch. Complex, sensitive or emotional conversations belong to people.
What drives the cost of an AI chatbot?
The model used and conversation volume (most models are billed by usage), preparing the content, integrations with existing systems and ongoing maintenance. There's a big difference between a simple setup with an off-the-shelf tool and a custom assistant connected to your ERP or CRM.
Can the bot work in more than one language?
Yes, modern models can answer fluently in many languages. Still, test each language separately; the best results come when your content also exists in that language.
In short
An AI assistant saves real time for your visitors and your team when it's grounded in the right content, has clear boundaries and hands over to a human easily. Set up badly, it becomes a risk that speaks wrongly on your brand's behalf. At Kasel, our conversational AI and chatbot projects cover the whole process, from defining the use case and preparing content to privacy settings and measurement. If you'd like to talk about whether your site would benefit from an assistant, get in touch.
