Using feedback to improve answers

Every answer carries a thumbs-up and a thumbs-down. They are the cheapest quality signal you have, and they point at your documentation rather than at the assistant.

What a rating records

A rating is stored with the question that produced it, the answer as the visitor saw it, the articles that answer cited, and the request's trace identifier. An optional comment can accompany it.

That is what makes a thumbs-down actionable rather than merely discouraging: you can read the exact exchange and open the exact article, without asking anyone to reproduce anything.

Two things worth knowing. The question and answer are sent by the visitor's browser at the moment they rate, so they record what the visitor saw rather than forming an independent audit trail — the trace identifier is the link back to what our systems actually produced. And payment card numbers, security codes, passwords and tokens are stripped from all three free-text fields before they are stored, on the same rules used for handoff transcripts, so a customer who pasted a card number into the chat has not left it in your ratings.

Why a thumbs-down is usually a documentation bug

When we look into a negative rating, the cause is nearly always one of three things, and only the first is about the assistant:

1. The search found the wrong article. Two of your articles cover similar ground and the weaker one ranked higher.
2. The article is out of date. The answer faithfully reflects what your documentation says, and your documentation is wrong.
3. The article does not exist. Your customers are asking something nobody has written down.

The second and third are documentation fixes, and they are permanent: correct the article and every future answer is correct. See Diagnosing a wrong answer.

What to do with them

Review negative ratings weekly at first. Most accounts find that a handful of documentation edits in the first month remove the majority of bad answers, because customers cluster tightly around the same gaps.

Send us anything that looks like a retrieval problem rather than a content problem, with the question attached. That is ours to fix, not yours.

What we do with ratings

We use them to find retrieval failures and to prioritise our own work. We do not use your customers' conversations to train models — see Subprocessors and where data goes.