Your agent cites its sources. Answers carry numbered references like [1], [2] that map to the articles the answer was built from, shown as clickable chips beneath the response.
Why this matters
An AI answer without a citation is unverifiable. A visitor can't tell a correct answer from a confident-sounding wrong one, and neither can you when you're reviewing what your agent said.
Citations also change the failure mode. When the agent can only assert what it can cite, its worst case becomes "I couldn't find that" rather than a plausible invention.
How numbering works
Sources are numbered once per turn across everything the agent consulted — including any follow-up searches it ran mid-answer. So [2] means the same article throughout a response.
Any citation the agent produces that doesn't correspond to a real retrieved source is stripped before the answer reaches the visitor. A reference you see is a reference that exists.
Source quality is visible
Citation chips carry their trust tier, so you can see at a glance whether an answer was built on a hand-written article or a scraped page.
Turning them off
You can hide citations via Show sources under widget behavior. We'd advise against it — visitors trust an answer they can check, and citations are how they check.
When there are no sources
If retrieval finds nothing relevant, the agent says it doesn't know rather than answering from general knowledge. That's the intended behaviour, and it's a signal: an unanswerable question becomes a knowledge gap for you to fill.