AI Chatbots for Your Website

How AI Chatbots Actually Work

Every AI chatbot follows the same core loop. A visitor types a question, the chatbot searches your knowledge base for relevant information, sends that information along with the question to an AI model, and returns the model's response to the visitor. This process is called Retrieval Augmented Generation, or RAG, and it is what separates a useful chatbot from one that makes things up.

The knowledge base is built from your own content. You upload documents, paste text, or point the system at your website, and the platform breaks that content into small chunks. Each chunk is converted into a mathematical representation called an embedding, which captures the meaning of the text in a way that allows fast similarity searching. When a visitor asks a question, the system finds the chunks most relevant to that question and includes them in the prompt sent to the AI model.

This approach means your chatbot answers from real information rather than general knowledge. A chatbot trained on your product documentation will quote your actual specs, pricing, and policies. A chatbot trained on your FAQ will give the same answers your support team gives. The AI model's job is to take the retrieved information and present it in a clear, conversational response, not to guess or invent answers.

What a Modern AI Chatbot Can Do

The simplest use case is answering questions, but AI chatbots handle much more than that. A well-configured chatbot can qualify sales leads by asking the right questions in conversation, collect contact information naturally without feeling like a form, book appointments by checking availability, and walk visitors through troubleshooting steps that would otherwise require a support email.

Knowledge base training is the foundation. Upload PDFs, Word documents, plain text, or crawl your entire website. The system indexes everything and makes it searchable. You can update the knowledge base at any time, and the chatbot immediately has access to the new information. See How to Train AI on Your Own Business Data for the complete guide on training approaches.

Multi-model support lets you pick the AI model that fits your needs. GPT-4.1-mini handles most support conversations well at low cost. Claude Sonnet provides more nuanced responses for complex topics. Reasoning models like GPT-5.2 work through multi-step problems. You can change models at any time without retraining your knowledge base.

Human handoff is critical for any business chatbot. When a conversation reaches a point where a human needs to step in, the chatbot transfers the full conversation history to your Live Operator Chat inbox. Your team sees everything the visitor said and the chatbot's responses, so the visitor never has to repeat themselves.

Voice input and output adds another layer. Visitors can speak their questions instead of typing, and the chatbot can respond with natural spoken audio using AI Voices. This is useful for accessibility, kiosk applications, and any situation where typing is inconvenient.

Email monitoring lets the chatbot watch a mailbox and respond to incoming emails using the same knowledge base it uses for live chat. Support emails get answered instantly using the same training data, and anything the chatbot cannot handle gets flagged for human review.

Choosing the Right AI Model for Your Chatbot

The AI model you choose affects response quality, speed, and cost. There is no single best model for every situation, and the right choice depends on what your chatbot needs to do.

GPT-4.1-mini is the default recommendation for most chatbots. It is fast, handles conversational tone well, follows instructions reliably, and costs 2-4 credits per response for typical support conversations. If you are building your first chatbot, start here.

Claude Sonnet excels at longer, more thoughtful responses. It handles nuance better than GPT-4.1-mini and tends to follow complex system instructions more precisely. It costs more per response but produces noticeably better output for topics that require careful explanation.

Reasoning models like GPT-5.2 work through problems step by step before answering. They are best for chatbots that need to do calculations, compare options, or work through decision trees. They cost more and respond slower, so they are not the right choice for simple FAQ-style chatbots.

Cheap models like GPT-5-nano cost very little per response and work well for simple, straightforward Q&A. If your chatbot mostly answers the same types of questions with short responses, a cheap model saves credits without much loss in quality. See our AI Models Guide for detailed comparisons.

What It Costs to Run an AI Chatbot

Chatbot costs break down into two parts: building the knowledge base and running conversations.

Training costs 3 credits per chunk of text when you upload documents or crawl your website. A typical business website with 50 pages might produce 200-400 chunks, costing 600-1,200 credits total (under $1.20). You only pay this once per piece of content, and you can add new content over time.

Conversation costs depend on the AI model and the length of the conversation. GPT-4.1-mini costs roughly 2-4 credits per response. A chatbot handling 100 conversations per day at an average of 4 messages each would cost about 800-1,600 credits daily ($0.80-$1.60). Claude and reasoning models cost 2-5x more per response but may resolve issues in fewer messages.

The platform charges credits for the AI model cost (with markup for platform keys) plus a small software fee per request. You can use your own API keys from OpenAI or Anthropic to reduce the AI model portion, or use the platform's built-in keys for convenience.

Setup and How-To Guides

Industry Use Cases

Comparisons and Decisions

Technical Deep Dives