Skip to content
Getting started6 min read

Try Test Chatbot before you go live

Ask your chatbot a real question from the dashboard drawer. It uses the same knowledge search visitors will get.

Browse topics

What it is

Test Chatbot is a private conversation with your own chatbot, from inside the dashboard. It uses the same knowledge search and the same answer model a real visitor gets — unless, on a paid plan, you picked a separate testing model on the AI model card — so what you see here is what they will see. You are not talking to a simplified preview.

It is a drawer, not a page. It slides in over whichever tab you are already on, so you can fix a knowledge source, reopen it, and re-ask without ever losing your place.

When you would use it

  • Right after you add your first knowledge source, to prove the chatbot can find it.
  • After every change to instructions or Call actions.
  • Before you activate the chatbot for real visitors.
  • When a visitor reports a bad answer and you want to reproduce it.

Test Chatbot is the single most useful habit in the product. Every minute here is a minute you do not spend explaining a wrong answer to a customer.

Where to find it

Open Dashboard → Chatbots → your chatbot, then select the Test Chatbot button in the chatbot header.

That button is present on every chatbot tab — Playground, Conversations, Analytics, Knowledge, Actions, Channels, Settings — so you never have to navigate anywhere first. There is no page URL for it and no menu item; the header button is how it opens.

Steps

  1. Open Dashboard → Chatbots and click the chatbot.
  2. Select Test Chatbot in the chatbot header. The drawer slides in from the side — you stay on the page you were on.
  3. Type a question that should be answerable from something you uploaded, and press Enter. (Shift+Enter gives you a new line instead of sending.)
  4. Read the reply properly. Ask yourself: is it correct, is it complete, and does it sound like your business?
  5. If sources are shown, open them and check the chatbot cited the right document, not merely a plausible one.
  6. If the answer is weak, fix the underlying knowledge or instruction, wait for training to finish, and ask again. Do not go live on a guess.
  7. Select New Chat when you want a clean session — a long thread carries context that can mask a problem a first-time visitor would hit.
  8. Close the drawer with the X, Esc, or the browser back button.

What to actually ask

A good test set is five questions, not one:

Ask thisBecause it catches
A question whose answer is verbatim in one documentWhether the source finished training at all
The same question in a customer's clumsy wordingWhether retrieval survives real-world phrasing
A question whose answer spans two documentsWhether it can combine sources instead of picking one
A question you deliberately have no answer forWhether your fallback message reads well
Something off-topic ("what's the weather?")Whether it stays in role instead of free-associating

The fourth one matters more than people expect. Most first-day complaints are about the wording of the "I don't know" reply, not about wrong answers. Fix it on Settings → Messages & language — see Chatbot messages and language.

What you will see

The drawer is a compact chat window with the chatbot's own avatar, colours, and welcome message.

  • An environment label. In the testing environment, the small analytics shown under each reply — confidence, timing, the sources that were used — are visible only to you as the owner. Visitors never see them. There is also a production-preview mode that shows the reply exactly as a visitor would get it.
  • A syncing notice. If knowledge is still processing, the drawer warns you that answers may be incomplete until training finishes. Believe it — retest afterwards.
  • Sources. When reference links are on and the chatbot actually found something, the sources are listed. See Sources and citations.
  • A not-ready reply if the chatbot is active but has never finished training. See Untrained mode.

It is real, not a sandbox

Three consequences people are surprised by:

  1. It spends message credits. Every turn is charged at the chatbot's testing model rate — the live model's, unless you picked a separate testing model. Testing thirty questions on the Free plan uses more than half your monthly allowance. See Message credits.
  2. It creates a real conversation. Once there is a reply to store, the thread appears in Conversations with the channel shown as Test chat. That is useful — you can reread your own tests — but it means the transcript is stored. Do not paste passwords, tokens, or customer personal data into it.
  3. Teammates can see it. Anyone with permission to view conversations on this chatbot can read your test threads.

Test Chatbot vs. what visitors get

Test ChatbotHosted page and widget
Who can open itYou and teammates with dashboard accessAnyone with the link, or any visitor to your site
Needs the chatbot to be activeNo — a draft chatbot answers hereYes — a draft returns a not-found page
Needs a widget token or allowed domainsNoThe widget does
Shows owner-only analyticsYesNo
Spends creditsYesYes

That first row is the important one. A working Test Chatbot does not mean your public page works. If Test Chatbot answers but the link 404s, the chatbot is still a draft — see Chatbot readiness.

Limits and plan notes

Test Chatbot is available on every plan, including Free, and does not require any integration. It is also the only way to try a chatbot on Free before sharing the hosted link, since the website widget needs a paid plan.

An empty knowledge base does not produce invented answers — the chatbot says it does not know rather than reaching for the open web.

Common problems

The drawer will not open.

You must be on a chatbot page for the header button to exist. If you are on the Chatbots list, open a chatbot first. If the button is there and nothing happens, refresh the page and try again.

Answers ignore the file I just uploaded.

Uploading is not the same as finished training. Open Knowledge and wait until the source shows as trained, then ask again. See Dataset status.

It answers, but the public link says the chatbot was not found.

The chatbot is still a draft. Test Chatbot works for owners regardless of activation. Activate it — see Activate or pause a chatbot.

I want to continue a real visitor's conversation.

Open the thread from Conversations and use Open full chat. That reopens the session in Test Chatbot so you can keep asking as the owner. It spends credits, and it is your session — the visitor does not see your messages. See Read a conversation.

The chatbot stopped replying mid-test.

Check the credits meter in the sidebar. Running out mid-session is the usual cause. See Out of credits.

Common questions

How do I open Test Chatbot?

Select the Test Chatbot button in the chatbot header. It is present on every chatbot tab and slides in over the page you are on.

Does Test Chatbot create a real conversation?

Yes. Once there is a reply to store, the thread appears under Conversations with the channel shown as Test chat, and teammates with access can read it. Do not paste secrets into it.

Does it spend message credits?

Yes. Every turn is charged at the chatbot's testing model rate — the live model's unless you picked a separate one. Testing thirty questions on the Free plan uses more than half the monthly allowance.

Test Chatbot works but my link says not found. Why?

Test Chatbot is the owner's private door and works even on a draft chatbot. Visitors need the chatbot activated. Activate it from Settings → Danger zone.

What should I actually ask?

Five things: something answered verbatim in one document, the same question in clumsy wording, something spanning two documents, something you have no answer for, and something off-topic.

Was this article helpful?

Ready to try it on your own content?

Create a free workspace, add a document, and ask the questions your team is tired of answering.