Chatbot pricing and feature pages use a vocabulary most business owners never needed until now: credits, grounding, deflection, outcomes, RAG. The same idea often goes by three names. Below are 30 terms in plain English, grouped by the question you are probably asking, and alphabetical within each group.
The 30 terms, A to Z: AI agent, answer bot, citation, conversation, crawl, credit, deflection, document upload, embedding, escalation, FAQ bot, follow-up question, grounding, guardrail, hallucination, knowledge base, large language model, lead capture, live chat, model-weighted pricing, outcome, prompt, prompt injection, rate limit, RAG, re-crawl, sitemap, system instructions, top-up, widget.
What kinds of chatbot are there?
There are three broad kinds on websites: bots that answer, bots that act, and people chatting live.
AI agent
A chatbot that can take actions as well as answer, such as looking up an order or booking a slot. More power means more to set up and more that can go wrong. AI agent vs chatbot covers the difference.
On TailGPT: TailGPT is an answer bot, not an agent. It cannot take an order, make a booking or look anything up outside your pages and documents.
Answer bot
A chatbot whose only job is to answer questions from a fixed set of material, usually a website and some documents.
On TailGPT: This is what TailGPT is. Each chatbot answers from one website's selected pages and your uploaded documents, and from nothing else.
FAQ bot
An answer bot tuned for short, direct answers to the questions people ask most, such as opening hours, fees or delivery times.
On TailGPT: FAQ is one of the four roles you can give a chatbot, alongside Support, Sales and Custom.
Live chat
A chat window answered by a person on your team in real time. It needs somebody watching it; an AI chatbot does not. AI chatbot or live chat compares the two.
On TailGPT: TailGPT has no live chat inbox. Every answer comes from the chatbot.
Widget
The chat bubble in the corner of a website that opens into a chat window, usually added by pasting a short piece of code into the site.
On TailGPT: One line of code per chatbot, pasted just before </body>. Name, logo, colour and greeting can change later without touching that code.
How does a chatbot learn about your business?
A chatbot learns about your business by reading the material you point it at, mostly your website and your files.
Crawl (crawler)
Reading a website page by page. The crawler is the program that visits each page and keeps its text. A good one opens pages the way a browser does, because many modern sites look empty to simpler tools.
On TailGPT: Each ticked page is opened in a real browser; menus, footers and cookie banners are dropped, and the Websites page shows any that failed and why.
Document upload
Adding files, such as a price list or a warranty, to what the chatbot reads, for knowledge that never made it onto a web page.
On TailGPT: PDF, Word (.docx) and text (.txt, .md) files, with no limit on how many. A scanned PDF with no text in it fails with a clear reason, and uploading a corrected file with the same name replaces the old one.
Embedding
A way of turning text into a list of numbers that captures its meaning, so a question can find a passage that means the same thing in different words: "Do you deliver to Nashik?" finding "we ship across Maharashtra".
On TailGPT: Pages and documents are made searchable by meaning, not just by matching words.
Knowledge base
The material a chatbot answers from: on most website chatbots, the pages and files you gave it.
On TailGPT: The selected pages of one website plus the documents you uploaded for it.
Re-crawl
Reading the website again so the chatbot picks up changes, such as a new price. Some tools do it on a schedule; others only when you ask.
On TailGPT: On demand, from a monthly allowance shared across your account. There is no automatic refresh, and documents are never touched by a re-crawl.
Sitemap
A list of a website's pages that the site publishes for search engines and other tools, which chatbot tools use to find pages.
On TailGPT: Your pages are listed from your site's own page list when the website is added, and you tick which to read.
How does a chatbot work out an answer?
A chatbot works out an answer by finding the relevant parts of its material and having a language model write a reply from them.
Citation (source)
A reference to the page or document an answer came from, shown to the visitor by some tools and kept for the owner by others. Why an answer should say where it came from explains why it matters.
On TailGPT: If an answer cites something that was not among the pages it read, the answer is thrown away. Visitors see the answer as plain text, without source links.
Follow-up question
A question that only makes sense after the one before, such as "and how much is that?". The bot needs some memory of the conversation to answer it. How much it can hold at once is sometimes called its context window.
On TailGPT: The bot is given the last three exchanges of that conversation, and nothing else.
Grounding
Keeping a chatbot's answers tied to a specific set of material, so it answers from your content rather than from whatever it picked up in training.
On TailGPT: Answers come only from your selected pages and uploaded documents. When they do not cover a question, the bot says plainly that the website does not cover it, instead of guessing.
Large language model
The engine under most AI chatbots: a program trained on huge amounts of text to write language. It writes fluent sentences about almost anything, which is why it needs grounding.
Prompt
The text a language model is given to respond to. On a website chatbot that is usually the visitor's question, the bot's instructions and the passages found in your content, put together.
RAG (retrieval-augmented generation)
Finding the passages that answer a question first (retrieval), then having the model write a reply from those passages (generation). It is how most website chatbots answer from your content instead of from memory. What is RAG? explains it plainly.
On TailGPT: TailGPT finds the relevant parts of your selected pages and documents and answers from those.
System instructions (role)
The standing instructions that tell a chatbot how to behave: its tone, how long to answer, what to point people to.
On TailGPT: You choose a role: Support, Sales, FAQ, or Custom with your own instructions. A role changes the manner, never what the bot is allowed to say. More in how to set your chatbot's role.
What keeps a chatbot's answers safe?
Rules around the model keep its answers safe: what it refuses, how fast it can be asked, and what it will not invent.
Guardrail
A rule that stops a chatbot answering something it should not, or stops an unsafe answer reaching the visitor.
On TailGPT: Attempts to give the bot orders, requests to write code and too many questions too fast all get the same sentence: "I can only answer questions about this website. Try asking about something on it." A refused question does not use one of the month's questions.
Hallucination
A confident answer that is not true, produced because the model writes plausible sentences whether or not it has the facts. What is an AI hallucination? defines it, and why AI chatbots make things up covers the fixes.
Prompt injection
Text written to make a chatbot ignore its rules and follow somebody else's instructions, either typed into the chat or hidden in a page or file the bot reads. What is prompt injection? covers both kinds.
On TailGPT: Instructions found on your pages or in your documents are marked and treated as something to read about, not obey.
Rate limit
A cap on how many questions can be asked in a short time, so one person cannot flood the bot or use up everybody's allowance.
On TailGPT: Each visitor has their own allowance per minute. Going over it gets the same refusal sentence as any other refused question.
What do the words on a pricing page mean?
The words on a pricing page name the unit you pay for, and the unit decides what a month buys. AI chatbot pricing: what you are actually paying for goes deeper.
Conversation
One visitor's chat with the bot, from first question to last. Some tools bill per conversation.
On TailGPT: TailGPT counts questions, not conversations. You can read every conversation your chatbot has had in Show chats.
Credit (message credit)
A unit some tools use instead of counting questions. Each answer spends a number of credits, and the number can vary: on some tools an answer from a premium model spends several credits where one from an economy model spends one. Message credits vs questions explains how the two ways of counting compare.
On TailGPT: There are no credits. One question is one question.
Model-weighted pricing
Pricing where an answer from a larger model costs more units than one from a smaller model, so the same plan buys fewer answers on the premium option.
On TailGPT: Not used. Every question costs the same one question from your plan.
Outcome (resolution)
A unit that charges when the bot achieves something, such as resolving a customer's question. What counts is defined by the vendor. It is common on customer-service agents sold with a helpdesk, where it is usually charged on top of per-seat fees. Per-resolution chatbot pricing, explained walks through how it adds up.
On TailGPT: TailGPT prices by questions a month; see pricing.
Top-up
Extra allowance bought on top of a plan when a month's allowance runs out, without changing plans.
On TailGPT: Questions only. They never expire, are used after the month's plan allowance, and need a live plan.
What do support and sales teams mean by these words?
Support and sales teams use these three words for what a chatbot does for them, and they are where tools differ most.
Deflection
A question the chatbot answered that would otherwise have become an email, a call or a ticket. Vendors measure it differently, so compare definitions before numbers.
On TailGPT: The dashboard lists the questions your chatbot could not answer, with how many times each was asked, and an answer rate. There is no export.
Escalation (hand-off)
Passing a conversation from the chatbot to a person on your team.
On TailGPT: There is no hand-off to a person. The Sales role points visitors to the next step your website already has, such as a contact page.
Lead capture
Collecting a visitor's name, email or phone number inside the chat so sales can follow up. Lead generation chatbot covers an answer-first approach.
On TailGPT: TailGPT does not collect contact details or show forms. It answers, and points to the contact routes your pages describe.
Which of these terms matter most when you compare tools?
The pricing unit matters most, then grounding, then what happens when the bot does not know. Ask each vendor what one unit buys, whether the bot can fall back on general knowledge, and what a visitor sees when your pages do not cover a question. For how TailGPT handles each, see how it works.