LanguageENES
ProductAug 18, 2026 · 4 min read

Why your AI assistant must cite every answer

Citations are not just a trust signal. They let your team find and fix answers that are wrong before a customer acts on them.

Mauro CarreraFounder
The assistant answering that the licence refund window is 14 days, citing the refund policy and the licence terms, with 1,842 of 2,050 answers cited
The assistant answering that the licence refund window is 14 days, citing the refund policy and the licence terms, with 1,842 of 2,050 answers cited

The risk everyone talks about with AI assistants is hallucination: an answer that is false. The risk that actually costs you is subtler. The assistant says something plausible, nobody verifies it, and a customer acts on it.

Citations are a maintenance tool first

When you see a citation under an answer, it looks like a trust signal. For your visitors, it is. For your team it is something more useful: a path back to the source.

If the window is wrong, the fix is in one of two named files rather than somewhere in your prompt.

When an answer is wrong and you can see it came from refund-policy.pdf page 4, you know exactly what to fix: which document, which page, which section. Without citations you are guessing at prompts and rewriting content blind.

A citation is only useful if it is the source that was used

Retrieving eight passages and listing all eight is not a citation, it is a bibliography. What you need is the passage the sentence came from.

What happens when the assistant has nothing to cite

There is no third option here, and the absence of one is the design.

InteractiveWhat an answer is allowed to do

The assistant says it does not know. If you have configured it, it offers to hand the conversation to a person. It does not improvise from general knowledge.

This matters because a plausible-sounding answer about your refund policy is worse than no answer at all. The first gets acted on. The second gets escalated.

Your refusals are a content backlog

The questions your assistant could not answer are collected in the Analytics tab under Questions. That list is written by your customers, which makes it better than any backlog you would have written yourself.

Sorted by how often each one was asked, so the top row is the paragraph worth writing this week.

Work down it by frequency. Fourteen people asking the same thing is a paragraph you are missing, not a bug.

The metric that tells you something is wrong

Check Answered from your knowledge base as a share of all answers. You will find it on the Overview tab, next to the answer count.

The second tile is the one to watch. Answers and citations should be the same number.

If that percentage drops, one of two things happened. Either people started asking about topics you have not written down yet, or an indexing job failed quietly.

A quiet indexing failure looks like a content problem

Both causes move the same number in the same direction. Check the document list for a failed job before you start writing new pages.

Either way, you notice before a customer complains.

Questions people ask

What if an answer is partially correct but cites the wrong source?

The citation shows you where the assistant found the information. If it came from a source that does not actually support the answer, that is a sign to rewrite that document or to index it differently.

Can I hide citations from my visitors?

You can configure what visitors see. Citations stay visible to your team in the analytics either way, so you can audit answers and fix errors.

What happens if my assistant answers without citing anything?

It will not. The assistant is built to cite every answer that comes from your knowledge base, or to say it does not know. There is no middle ground.

Not quite. A link points at a page. A citation names the document and the passage inside it that the sentence came from, which is what makes it useful for repair.

Start by opening the Questions list, taking the top three, and writing the paragraphs your customers have already asked for.

Taggedcitationstrustsupport
Written byMauro CarreraBuilds WelcomeAI. Spends most days reading similarity scores that should have been higher.

Keep reading