In short
Only if it answers questions better than the page can, and the volume of those questions justifies the cost of doing it well. Chatbots help on sites with many repetitive support questions, complex products where a guided conversation beats browsing, and after-hours lead capture. They hurt on small sites where a bubble covers the content and a bot that cannot answer sends visitors away. An LLM-based bot grounded in your own content can be genuinely useful; an ungrounded one confidently invents policies. Fix the FAQ and the content first, then add a bot if the questions keep coming.
What a chatbot is competing with
A visitor with a question has options: read the page, search the site, look at the FAQ, email, call, or ask the bot. The bot wins only when it is faster or better than those. On a site whose pages already answer the common questions clearly, the bot's ceiling is low and its floor, a bubble that covers the pricing table and a bot that says "I didn't understand that", is well below the alternatives.
So the first question is not whether to add a chatbot. It is whether the site's content is doing its job. A FAQ that answers the twenty questions customers actually ask, and pages that state prices and processes plainly, resolve most of what a bot would be asked, and they also rank and get cited, which the bot does not.
Where a chatbot genuinely helps
High-volume repetitive support. Order status, password resets, opening hours, return policies. If the same fifty questions arrive hundreds of times a week, a bot that answers them accurately saves real support cost and gives customers an instant answer.
Complex products with guided paths. Choosing the right plan, configuring an option, finding a part. A conversation that narrows down the answer can beat a visitor reading six pages.
After-hours lead capture. A bot that takes a name, a question and a preferred time, and hands off to a person in the morning, captures enquiries that a contact form loses.
Large content libraries. Documentation and knowledge bases where search is weak; an LLM bot grounded in the docs can find the answer and cite the page.
Businesses already handling chat. If people are answering live chat anyway, a bot that triages, gathers details, and hands off is an efficiency, not a replacement.
Where it hurts
Small sites with modest traffic. A bubble on every page, a greeting that pops up uninvited, and a bot that answers three questions and fails the fourth. The visitor learns the business is hard to talk to.
As a substitute for clear pages. If the answer is only available by asking the bot, the site has hidden its content behind a conversation, and search engines and AI assistants never see it.
As a substitute for people. Visitors who wanted a human and got a bot that cannot escalate leave angry, and say so in reviews.
Mobile. A chat widget on a phone can cover a third of the screen and the primary button. Many implementations are worse than the problem they solve on the majority of visits.
Scripted bots and AI bots
Scripted (rule-based) bots follow decision trees. Predictable, cheap, safe, and limited to what someone wrote. Right for narrow tasks: booking, triage, a handful of FAQs.
LLM bots generate answers. Grounded in your own content through retrieval, they can answer a wide range of questions in natural language and cite the source. Ungrounded, they invent: policies you do not have, prices you do not charge, promises a court may hold you to, as one airline discovered. An LLM bot needs retrieval over your own content, guardrails on what it will discuss, a clear statement that it is a bot, an escalation path to a person, and logging of every conversation so you can see what it said.
Built that way, an LLM bot is a real improvement over the scripted kind. Built as a raw model with a system prompt, it is a liability with a chat window.
What it costs to run well
The widget is cheap. The work is not:
- Content to ground it, kept current. A bot answering from last year's pricing page is worse than no bot.
- Conversation review. Someone reads transcripts weekly to find what the bot got wrong and what people ask that the site does not answer. This is the most valuable output of a chatbot and the step most teams skip.
- Escalation. A human reachable from the conversation, with the transcript, during defined hours, and a clear message outside them.
- Per-message costs for LLM bots, which scale with traffic and are worth estimating before launch.
- Privacy. Conversations are personal data; the widget vendor is a processor; the policy needs updating.
A bot with none of this is a bubble.
The decision checklist
- What are the twenty questions people actually ask? Are they answered on the site? If not, fix that first.
- How many questions arrive per week? Below a few dozen, a bot is not worth its upkeep.
- Who reviews conversations, and when?
- Where does an unanswered question go, and how fast?
- Will the widget be hidden or minimal on mobile and on pages with a primary action?
- For an LLM bot: what content grounds it, what is off-limits, and what does it say when unsure?
If those have good answers, add the bot. If they do not, the site's clearest pages will serve visitors, search engines and AI assistants better than a conversation ever could. If you would like help deciding, or building one that is grounded and supervised, book a call.
Common questions
Do chatbots increase website conversions?
Sometimes, on sites with complex products, high enquiry volumes or after-hours lead capture, where a bot answers questions the pages do not and hands off to a person. On small sites with clear content, a chat bubble more often covers the call to action and answers poorly, and conversions fall. Test it, and measure conversions rather than chat volume.
Are AI chatbots safe to put on a business website?
Only when grounded in your own content through retrieval, restricted to relevant topics, clearly labelled as a bot, able to escalate to a person, and logged so someone reviews what it said. An unconstrained model will confidently invent policies and prices, which has led to legal liability for at least one company. Built with those controls, an AI bot is a genuine improvement over scripted ones.
How much does a website chatbot cost?
Scripted bots run from free to a few hundred dollars a month for the software; LLM bots add per-message costs that scale with traffic. The larger cost is upkeep: keeping the grounding content current, reviewing transcripts weekly, and staffing the escalation path. A bot without those is cheap and worse than none.
What is the difference between a chatbot and live chat?
Live chat connects the visitor to a person in real time; a chatbot answers automatically from scripts or a language model. Many sites combine them, with the bot handling routine questions and gathering details before handing off to a human. The combination works when the handoff is fast and the transcript travels with it.
Should a small business website have a chatbot?
Usually not as a first step. A small business does better with pages that state prices, processes and hours plainly and a FAQ covering the questions customers actually ask, which also help search rankings. Add a bot only if enquiries keep arriving in volume after that, and keep it minimal on mobile.
