Every reply the Chatixy assistant sends is checked, the moment it is saved, for one thing: did the visitor ask something the assistant could not answer? When the answer is yes, the reply is stamped as an answer gap. The stamped replies are collected on one page, Answer gaps, under the Knowledge Base. It is the most practical report in the dashboard, because each row on it is a question a real visitor asked and your content did not cover.
This article explains what puts a reply on that list, what you see for each row, and how to clear it so the same question gets a proper answer next time.
Three ways a reply becomes a gap
A reply is stamped for one of three reasons, and the page tells you which.
Nothing in your knowledge base covered it. When a visitor asks a real question and the search across your website pages, documents, texts and Q&A pairs finds nothing to answer from, the reply is marked as having no knowledge behind it. This signal reads what the assistant had to work with, not the words it used, so it behaves the same in every language the widget speaks. It is also the most useful reason to see, because it names the fix directly: the answer is missing from your content.
The assistant said it could not help. Replies that decline in the first person, such as "I don't have that information" or "our team will confirm that for you", are also caught. This check reads the wording, and it only reads English. It is kept because it is the only signal older conversations can carry.
The visitor pressed "not helpful". A thumbs-down under a reply puts it on the list too, with the visitor's own explanation if they typed one. Of the three, this is the strongest signal: nobody guessed, the person who asked said the answer did not work. If the visitor changes their mind and presses thumbs-up, the row disappears again.
What is kept off the list
A report full of false alarms stops being read, so several kinds of reply are deliberately left out.
- Small talk. "Hi?", "thanks" or "ok" are not unmet needs. A message counts as a question only when it has a question mark or is long enough to be one, and very short greetings are filtered out even when they end in a question mark.
- A no that comes with an alternative. If a visitor asks for a product your shop does not list, and the assistant says so and links to a real product from your own site instead, that is a good answer, not a gap.
- A no that still answers. "Our standard delivery normally takes 2-7 working days. I don't have your postcode, so I can't say exactly when it would arrive." gives the fact that was asked for and is honest about the rest. When the reply states the kind of fact the question asked for, such as a price, an opening time, a delivery time, an address or a contact detail, it is not counted.
- Your own tests. Conversations from "Test it live" in the dashboard are you checking your assistant, not a visitor it failed, so they never appear.
- Outages. If the AI service itself fails and the visitor sees the "having trouble responding" message, that is a service problem. No answer you write would fix it, so it does not land in your list.
What each row shows
Each row has the question the visitor asked, the first part of the reply the assistant gave, the conversation it came from, the assistant it belongs to if you run several, and why it was flagged. For a thumbs-down, the visitor's own words are shown when they left some. You can narrow the list to the replies visitors marked as not helpful.
Closing a gap
A row stays on the list until you close it, and there are two ways to do that.
Improve the answer. This opens the question and the reply side by side. You edit the question if needed, write the answer you want the assistant to give, and save. The answer is stored as a Q&A pair for that assistant, which it uses for similar questions from then on, and the row leaves the list. The count in the sidebar and on the Knowledge Health card drops at the same moment.
Dismiss it. Some flagged replies are fine. "We don't have that in red, but we do in blue" contains a decline and is still a good answer. Dismissing records that you judged it a false alarm, and the row does not come back.
Both actions record who closed the row and when.
Let it draft the answer for you
The hardest part of improving an answer is the empty box. On the Answer gaps page, the improve dialog has a Suggest answer button. It searches your knowledge base for the question, reads the turns of the conversation around it, so that "do you have it in red?" knows what "it" is, and adds your company details such as your phone number, hours and address. From those it writes a short draft in the visitor's language.
The draft is told to use only what you have given it. Where your content does not cover something, it leaves a placeholder in square brackets, for example [delivery time], for you to fill in, rather than making a detail up. It also shows the sources it drew on. Nothing is saved until you edit it and press save, exactly as if you had written it yourself.
A few things to know before you rely on it:
- Suggestions are made only when you click, never when the page loads.
- Each workspace has a daily allowance of 200 suggestions. When it is used up, the dialog says so and you can still write the answer by hand.
- Drafting and saving an answer need permission to manage the knowledge base. Seeing the report needs permission to view analytics.
A weekly habit that pays off
The list gets shorter only when someone works through it. A sensible routine is to open Answer gaps once a week and work through the rows with no knowledge behind them first, because each of those is a question your content does not answer yet. Then look at the thumbs-down rows and read what the visitors wrote.
If you have not opened it yet, go to Knowledge Base in your Chatixy dashboard and start with the newest row.