Conversation feedback

What it is

Feedback is the visitor's own verdict on a reply: a thumbs up (Helpful) or a thumbs down (Not Helpful). It appears as a chip on the message in the transcript and as a filter on the conversations list.

It is not a public star rating, it is not an average score shown anywhere to customers, and it is not the same as your team's Reviewed flag. Feedback is what the visitor thought; Reviewed is whether you have dealt with it.

Most replies never get either thumb, and that is normal. Visitors only rate when they feel strongly, which is exactly what makes the ones you do get worth reading.

When you would use it

After go-live, on a weekly rhythm. Thumbs-down replies are the shortest path to knowing what is actually wrong with your chatbot — better than guessing, and better than reading everything.

Where to find it

Open Dashboard → Chatbots → your chatbot → Conversations, and use the Feedback filter. Individual thumbs appear on messages in the transcript.

Your chatbot's negative feedback is also summarised on Analytics → Needs attention, alongside unanswered questions. See The chatbot Analytics tab.

Steps

1. Open the chatbot → Conversations.
2. Set the Feedback filter to Negative. Optionally add Review status: Needs review so you only see what nobody has handled.
3. Open a thread with View details.
4. Find the rated message and read the whole exchange around it — the reply before and the message after usually explain the rating better than the rated message alone.
5. Open the sources on that reply. Did the chatbot cite the right document? Did it cite anything at all?
6. Diagnose it as one of the four causes below, and fix it in the matching place.
7. Re-ask the same question in Test Chatbot after any knowledge change has finished training.
8. Mark the conversation Reviewed.

The four reasons people press thumbs-down

Diagnosing correctly is the whole job, because each one has a different fix.

1. The answer was wrong or missing. The chatbot did not know, or cited something outdated. → Fix the knowledge. Add the missing document, or remove the stale one. See Knowledge overview and Remove knowledge.

2. The answer was correct but useless. Technically accurate, unhelpfully vague, or four paragraphs when one would do. → Adjust length on Chatbot AI settings, or add an instruction about how to answer this class of question. See Write instructions.

3. The tone was wrong. Too stiff, too chirpy, or it said something you would never say. → Instructions. This is exactly what they are for.

4. They did not want an answer at all. They wanted a person, a refund, or a callback, and a paragraph of documentation felt like being fobbed off. → No amount of knowledge fixes this one. Give them a route out — see Handoff to a human or Notify your team.

Cause 4 is the one people miss most often. If a thumbs-down looks completely correct to you, it is usually this.

Reading feedback honestly

A few cautions before you rebuild your knowledge base around three thumbs:

• Volume is low by nature. A handful of ratings out of hundreds of conversations is normal. Do not treat two thumbs-down as a trend.
• Negativity bias is real. Frustrated people rate more often than satisfied ones. A poor ratio does not automatically mean a poor chatbot.
• Cross-check with the harder numbers. Answered rate and average confidence, both on the conversations list, cover every conversation rather than the few that were rated. Use feedback to find examples; use those numbers to judge scale.
• Not every surface collects thumbs. A message with no chip means nobody rated it — never a hidden neutral.

Feedback vs. Reviewed vs. confidence

	Who sets it	What it means
Feedback	The visitor	Their opinion of one reply
Reviewed	Your team	Somebody has looked at this thread
Confidence	The system	How well-grounded the answer was in your knowledge

They disagree usefully. A high-confidence answer with a thumbs-down is almost always cause 3 or 4 — the chatbot found the right document and the visitor still did not want it. A low-confidence answer with no rating at all is a gap nobody bothered to complain about, which is worth finding before someone does.

What you will see

On the list, a Feedback filter with All, Positive, Negative, and No feedback. In the transcript, a small chip on rated messages alongside the status and timing chips.

Limits and plan notes

Reading feedback needs permission to view conversations, and it does not spend credits. Available on every plan.

There is no control on the transcript for clearing or overriding a visitor's rating — it is their record of what they thought. There is also no way to reply to a thumbs-down as the chatbot; improving the answer for the next person is the intended response.

Common problems

No conversation has any feedback.

Visitors have not rated anything yet, which is common early and on low-traffic sites. Use Review status for your own queue, and Answer status: Not answered to find gaps without waiting for ratings.

A thumbs-down looks completely correct to me.

Then it is probably cause 4 — they wanted a person or an outcome, not information. Check whether the chatbot offers any route to a human.

Lots of thumbs-down on the same question.

Fix that one question properly rather than tuning settings. Find the document that should answer it, confirm it is trained, and re-ask in Test Chatbot until the answer is right.

Can I delete a visitor's thumbs-down?

No. Spend the effort on the answer instead.

Is Mark reviewed the same as a thumbs up?

No. Reviewed is your flag; the thumb is theirs. Marking something reviewed does not clear or contradict a negative rating.