← Help Centre

Testing · 5 min read ·

How to Test Your AI Chatbot Before Launch

Use this guide after adding your business knowledge and before inviting customers to use your chatbot. You will check the facts, the conversation, and the live widget experience.

1. Prepare your test questions and approved answers

Use a workspace admin account. Confirm your website profile preparation and document processing have finished, and save any changes to your assistant settings. If you have not completed setup, start with the installation guide.

Create a small checklist with four columns: customer question, expected facts, actual reply, and pass or needs work. Use current business information as the reference.

  • Common question: a product detail, opening hour, or published policy.
  • Buying question: which service or plan fits a stated need.
  • Ambiguous question: “How much does it cost?” without naming a product.
  • Unknown detail: a discount or exception you have not published.
  • Human request: “Can I speak to someone?”
  • Account action: a cancellation, refund, or change that requires your team.

Write expected facts rather than one exact sentence. A correctly phrased answer can use different words.

2. Run a fresh dashboard test

  1. Open Test Your Agent.
  2. Select New Test to start a fresh conversation.
  3. Ask the first question exactly as a customer might type it.
  4. Compare the reply with your approved facts. Check any prices, conditions, dates, and links.
  5. Ask a natural follow-up, then rephrase the original question in another fresh test.
  6. Repeat for each case on your checklist.

For example, ask “Do you deliver to my area?” An assistant needs the relevant location before giving a location-specific answer. After supplying the location, check whether the reply follows your real coverage policy.

Mark a reply as needing work if it states an unsupported fact, misses a material condition, or claims to have completed an action it cannot perform. Do not judge readiness only by a friendly tone.

3. Check unknowns and human help

Ask a question whose answer is absent from your sources. The reply should make the uncertainty clear and offer a useful next step. Instructions can guide this behavior, but you need to test it.

Request a person and review your configured handoff process. Confirm that the visitor understands what happens next and that your team can locate and follow up on the conversation. Check the wording outside support hours too.

Use test details for this exercise. If an account action requires a person, the assistant should explain that route rather than say the action is done.

4. Fix failures and repeat the affected tests

  • Wrong or missing fact: correct the website profile source, document, or Q&A pair.
  • Conflicting answers: find and update superseded content across your sources.
  • Unhelpful tone or length: adjust Agent Settings → Identity → Agent Description and select Save Changes.
  • Unclear next step: supply the approved process and verified destination link.

Follow the answer improvement guide for detailed source-update steps. Wait for processing to complete and begin a New Test after changes. Retest the failed case and a related question to check that the correction did not create a new contradiction.

5. Verify the widget on your live website

After installing the dashboard’s embed snippet, refresh your website and start a new chat. Test Your Agent checks the assistant, while this step checks what your visitors actually experience.

  1. Open a page that includes the widget on a desktop browser and a phone.
  2. Check that the chat button is visible and does not obscure navigation, forms, or checkout controls.
  3. Open and close the panel. Check that text and controls fit the screen.
  4. Ask a known question and follow-up. Open a supplied link to confirm the destination.
  5. Request human help and confirm the visitor instructions and team follow-up.

If the button is missing, check that the correct workspace snippet is installed on the published page and review any browser errors or site restrictions. The setup troubleshooting section covers the next checks.

6. Decide whether you are ready to launch

Launch when your key questions receive accurate, usable replies, important conditions are preserved, unknowns are acknowledged, and human help works as intended. Resolve material errors before launch.

Keep the checklist and assign someone to review early conversations. Repeat the relevant tests after changing prices, policies, sources, or instructions. A launch check is a starting point for ongoing maintenance.

If you need help, contact support with the question, actual answer, expected facts, and relevant source or error. Remove private customer details from the example.