Skip to main content

​
Assistant Response Issues

If your AI assistant does not respond as expected, start by identifying what the visitor sees. A response that never appears needs a different investigation from an answer that appears but contains incorrect information. This guide covers common symptoms, practical checks, and the information to collect when you need support.

​
Identify the issue

A temporary response failure does not necessarily mean your knowledge files are missing or your assistant configuration has changed. Check whether the problem is reproducible before making broad changes.

​
Start with a quick diagnostic

  1. Open your AI Experience in a new browser tab.
  2. Start a new conversation if that option is available.
  3. Send a short, straightforward question that your assistant should be able to answer.
  4. Wait for the response to finish before sending another message.
  5. Repeat once with a different question.
  6. Note whether the problem occurs consistently, intermittently, or only in a particular conversation.
For a public experience, you can also test the visitor-facing page in a private browser window. For private or invite-only experiences, use an account with the appropriate access.
Test with a simple question first, such as “What can you help me with?” Then test a specific question based on your uploaded knowledge. This helps distinguish a general response failure from a knowledge-retrieval problem.

​
No response or an error

If sending a message produces no answer, or the page displays an error:
  1. Check your internet connection and reload the page once.
  2. Confirm that you are on the intended AI Experience URL.
  3. Verify that your session or visitor access is still valid.
  4. Check whether the experience is available and whether your subscription or message allowance needs attention.
  5. Try a new conversation and a short message.
  6. If possible, test in another browser or device.
If the issue persists, record the exact error text and approximate time. Avoid repeatedly resending the same message, as this can make troubleshooting harder.
Do not paste API keys, passwords, payment details, or sensitive visitor information into a support report or test conversation.
Related guides: Message Usage and Limits and Sign-In Problems.

​
Slow or interrupted responses

AI responses may take longer when a question is complex, the answer requires more context, or the connection is unstable. A persistent typing indicator, however, may indicate that the response did not finish correctly. Try the following:
  • Wait briefly to see whether the response completes.
  • Test with a shorter question.
  • Check whether the browser tab or network connection was interrupted.
  • Start a new conversation to see whether the behavior is isolated to one chat.
  • Compare desktop and mobile behavior if the issue appears device-specific.

​
Streaming interruptions

Some interfaces display an answer progressively as it is generated. If text appears and then freezes, note whether:
  • The typing indicator remains visible.
  • The answer stops at the same point when retried.
  • The message input becomes unresponsive.
  • The conversation jumps or scrolls unexpectedly.
  • The response appears correctly after refreshing the page.
These observations can help support distinguish a response-generation issue from a chat-interface issue.

​
Incomplete answers

An answer may end abruptly, omit important steps, or fail to address the full question. First, ask a focused follow-up such as “Please continue from the last point.” If this works, the original request may have been too broad. If incomplete answers occur repeatedly:
  1. Break long, multi-part requests into smaller questions.
  2. Review the assistant’s system instructions for conflicting or overly restrictive directions.
  3. Check whether the relevant source material is complete and clearly organized.
  4. Test the same question in a fresh conversation.
  5. Record a reproducible example for support.
For configuration guidance, see Testing Your Assistant.

​
Incorrect or unsupported answers

If the assistant responds but gives an inaccurate answer, distinguish among three possibilities: For example, an assistant for a boat broker should not claim that a particular boat is currently available if it does not have access to live inventory. Instead, it should explain its knowledge limits and direct the visitor to the broker for current listings. Likewise, an educational insurance assistant should not invent policy terms, personalized eligibility, or coverage guarantees.
For legal, medical, financial, insurance, or other high-stakes topics, verify important information independently and use an appropriate qualified professional for decisions. An AI response should not be treated as a substitute for professional advice.
See Knowledge Retrieval Problems for a deeper investigation of missing or mismatched source content.

​
Unexpected behavior

If your assistant changes tone, ignores boundaries, uses the wrong language, or makes an inappropriate recommendation, review its configuration.

​
Check the instructions

Look for:
  • Conflicting directions about the assistant’s purpose or audience.
  • Unclear rules for when to answer versus when to recommend human follow-up.
  • Instructions that unintentionally encourage invented information.
  • Missing guidance on responding in the visitor’s language.
  • Overly aggressive promotional instructions that interrupt helpful answers.
Make one targeted change at a time, then retest with the same question. This makes it easier to identify what improved the behavior.

​
Check the knowledge

Knowledge files provide reference information; system instructions define how the assistant should behave. Updating one does not always solve a problem caused by the other. For instance, adding a document about available services may improve factual coverage, while instructions are better suited to controlling tone, uncertainty, and handoff behavior.

​
Access and browser checks

If the assistant works for the owner but not for visitors, check the visitor experience separately.
  • Public experience: Test from a signed-out or private browser session.
  • Private or invite-only experience: Confirm the visitor has the required access.
  • Guest chat: Verify the expected visitor journey, including any registration step.
  • Custom domain: Test both the configured domain and the expected destination URL, where applicable.
  • Mobile-only issue: Try a second mobile browser and record the device type.
See Public and Private Access, Guest Chat and Registration, and Custom Domains and DNS.

​
A repeatable troubleshooting workflow

Use this sequence before changing multiple settings:
  1. Reproduce: Capture the exact question and observed result.
  2. Isolate: Determine whether it happens in one chat, across chats, or only for certain visitors.
  3. Classify: Decide whether the issue concerns access, generation, knowledge, instructions, or the interface.
  4. Change one variable: Update only the relevant setting or source.
  5. Retest: Repeat the original question and compare the result.
  6. Document: Keep the successful fix and any remaining limitations.
A successful test should confirm not only that the assistant produces an answer, but also that the answer is useful, accurate within its knowledge, and consistent with the intended visitor experience.

​
When to contact support

If the problem continues after basic checks, prepare the following information: Screenshots or a short screen recording may help when the problem involves loading indicators, scrolling, or the interface. Remove or obscure personal and confidential information before sharing them. For assistance, see Report a Bug or Contact Support.

​
Before you finish

  • I tested a simple question in a new conversation.
  • I checked whether the issue is consistent or intermittent.
  • I verified visitor access and the correct website URL.
  • I reviewed relevant instructions or knowledge only when appropriate.
  • I changed one variable at a time and retested.
  • I recorded the error and reproduction steps if support is needed.