AI chatbot for your website, ready in minutes

Classic

The default. A quiet rounded panel that sits on almost any site without argument.

Calm

Softer contrast and roomier spacing, for content-heavy sites where chat should not shout.

Contrast

Heavier weights and firmer edges. Stays readable on busy or image-led pages.

Executive

Restrained and formal - the one to pick for finance, legal and B2B sites.

Glass

A translucent panel over a blurred backdrop. Best on sites with strong photography.

Hero

A larger opening panel that gives the welcome message and starter prompts real space.

Persona

Built around a named assistant with an avatar, for when the AI should feel like someone.

Portal

Structured and dense, for help centres where people arrive already looking for an answer.

Warm

Rounder shapes and a friendlier palette, suited to retail and hospitality.

Your policies, explained in a conversation.

Give visitors a useful answer from the content you maintain. When the knowledge is missing, ZChat can explain the gap and route the conversation to your team.

Illustrative conversation: Can I return a gift without a receipt? Our gift-return policy allows an exchange with the order number. Do you have it handy?

• Find relevant business content
• Answer using that knowledge
• Keep the conversation in your inbox

Connect the answer to your business systems.

Configure an order lookup or your own API action so the AI can use current information. You choose the connected tools and the actions it can run.

Illustrative conversation: Where is my order #1042? Your order has shipped. The carrier currently estimates delivery on Thursday. Here is your tracking link.

• Collect the information your action needs
• Look up the order in your connected system
• Return the result in the same chat

Let your team take it from here.

A customer can ask for a person without losing the conversation. Your teammate receives the transcript; offline requests stay in the queue.

Illustrative conversation: I need help with a damaged delivery. Can I talk to someone? Hi, I’m Sam from support. I’ve read your conversation. Could you share a photo of the damaged item?

• Recognize the request for a person
• Keep the transcript with the handoff
• Reply from the shared inbox

It says “I don’t know” instead of inventing an answer.

The answers it does give are grounded in pages you wrote — and you can see which ones.

A language model always produces fluent text. Ask a support bot something your site never covered and most will still answer — confidently, in your brand voice, and wrong. The customer believes it, acts on it, and the ticket you get back is worse than the one you avoided.

ZChat scores every candidate article against the question and weights each word by how rare it is in your own content, so words that appear on every page cannot decide the match. Below a minimum relevance nothing is sent to the model at all, and the AI says it could not find the answer on your site.

Asking for a human is not a dead end at 2 AM.

Nobody leaves because nobody answered, and the transcript is waiting when your team arrives.

On most tools the bot goes silent the moment someone asks for a person. If your team is offline — evenings, weekends, the middle of a launch — the visitor is left typing into a window that has stopped replying, and they leave.

ZChat keeps their place in the queue and lets the AI carry on, saying once that nobody is available and that their request still stands. If an agent has actually replied, the AI stays out of the way for good.

Fix a wrong answer once and it stays fixed.

Your assistant improves as a by-product of ordinary support work.

Correcting a bad reply usually helps exactly one customer. The next person asking the same question gets the same mistake, and improving the knowledge base is a separate chore that never reaches the top of the list.

Edit the reply in your inbox, leave the box ticked, and the correction becomes a knowledge base article the AI answers from next time. The original wrong answer is kept, so anyone can review whether the correction was right.

You test the change. Not your customers.

Regressions are found by you, on a Tuesday, instead of by a customer on a Saturday.

Changing an instruction or switching models is a one-line edit with no visible blast radius. The reply that improves is the one you were looking at; the three that quietly got worse are the ones you were not. Most tools ship that change straight to customers and let them find the regression for you.

Run multi-turn conversations in a sandbox and compare a candidate model or instruction change against your current one before you save it.

“Why doesn’t it know about that?” has an answer in the product.

A gap is a task with a name, not a mystery.

When a bot misses something, most tools give you nothing to look at. A page that failed to crawl and a page that was never crawled look identical from the outside: silence.

Every source is listed with what actually happened to it — indexed, skipped or failed, and why. Alongside it you get the questions nobody could answer, the replies visitors marked unhelpful, and the confidence and sentiment behind each conversation.

Your AI key, or ours. Your bill either way is predictable.

You can forecast the cost, and you choose the data path.

Usage pricing makes your invoice a function of your busiest month, and per-resolution fees charge you more precisely when the product is working. Managed-only AI also means every conversation runs through somebody else’s account whether that suits you or not.

Run on ZChat-managed AI, or connect your own AI provider key and pay the provider directly. Either way you are billed per agent seat — no usage credits, no per-resolution charge.