The question behind the question
A prospective customer asks your assistant whether your warranty covers accidental damage. It answers: "Yes, accidental damage is covered for the first 24 months."
Now watch what happens in their head. It is not "great, thanks." It is is that true? They have read enough about AI to know that a confident sentence and a correct sentence look identical. So they do one of three things: believe it and get annoyed later if it was wrong, ignore it and email you anyway, or leave.
All three outcomes are bad, and none of them are fixed by making the model better. A more accurate assistant still produces sentences that look exactly like the inaccurate ones. What is missing is not accuracy. It is evidence.
What grounding actually buys you
The standard fix is retrieval: before answering, search the customer's own documents and put the relevant passages in front of the model along with the question. The model answers from real source material rather than from whatever it absorbed in training. Building a knowledge-base agent with RAG covers the mechanics.
Grounding genuinely reduces the failure rate. It does not do anything for the reader, because the reader cannot see it. From the outside, a grounded answer and an ungrounded one are the same block of text. All the work is invisible at precisely the moment it needs to be visible.
That is the gap a citation badge closes: a small marker under the answer naming the document it came from. It converts "trust me" into "check me," which is a much easier thing to ask of a stranger.
The badge has to be true, and that is harder than it looks
Here is the trap, and we walked into it.
Retrieval and use are two different events. The search returns a handful of candidate passages — typically three to five. The model then reads them and writes an answer, and in doing so it may lean entirely on one, blend two, or ignore all of them because none were relevant and it answered from the conversation instead.
The easy implementation cites everything the search returned. It is one line of code and it is wrong. If the answer came from your shipping policy and the badge also names your returns policy and your terms of service, you have taught the reader that the badge means "documents we looked at," which is a fact about your infrastructure and not about the answer. Worse, someone who follows the citation to verify a claim and cannot find it there concludes the whole thing is decorative. A citation that does not survive being checked is worse than no citation, because it converts a neutral reader into a suspicious one.
So the badge has to report what the model actually read, not everything the search offered up. That means reporting the passages that were genuinely placed in front of the model for that answer, rather than the full candidate list the index produced — and it means the badge sometimes names one source where the search returned four. That is the correct outcome. The number in the badge is not a measure of effort.
| Approach | What the badge shows | What the reader learns |
|---|---|---|
| No citation | Nothing | Whether to trust you, based on vibes |
| Cite the candidate list | Everything the search turned up | That citations are noise |
| Cite what was read | The source the answer rests on | Where to verify, in one click |
The other half: saying nothing
The most useful thing a grounded assistant does is refuse.
If someone asks about a policy you have never written down, there is no honest citation available, and the assistant has two options. It can synthesise something plausible from adjacent documents and general knowledge — fluent, unattributable, and occasionally a commitment you never made. Or it can say it does not have that information and offer to get a person involved.
The second is correct, and it is the behaviour most teams accidentally tune out of their assistant. The pressure runs one way: a prompt that says "be helpful" and a model that is very good at sounding helpful will fill silence rather than admit a gap. You have to ask for abstention explicitly, and then you have to test that you still have it, because it is the first thing to break when someone edits the prompt for an unrelated reason. Testing for behaviour rather than output shape is its own discipline — testing AI assistant behaviour covers how we grade it.
A well-behaved gap looks like this: "I do not have anything on accidental damage in the documents I can see. I can pass this to the team with your question — what is the best email to reach you?" Nobody is delighted by that answer. Everybody trusts it more than a fabricated yes.
What citations change about how buyers behave
Three things, consistently.
High-stakes questions get asked
People do not ask an AI about refund eligibility, contract terms, or medical restrictions if they expect a guess. Visible sourcing moves those questions from "email the company" to "ask the assistant," which is exactly the traffic you wanted the assistant to absorb in the first place.
Objections surface earlier
When a buyer can see that your answer comes from a policy document dated eighteen months ago, they say so. That is a gift. The alternative is that they silently discount everything the assistant tells them and you never find out why the conversation went cold.
Your own team starts using it
Once the badge is trustworthy, internal users treat the assistant as a search index over the document set rather than as a novelty. That is usually the first real sign the knowledge base is working.
Practical notes
- Keep the badge quiet. It sits under the answer, small, one tap to open the source. It is a reference, not a banner. An answer that has to defend itself visually reads as less confident, not more.
- Cite the document, and the section if you can. "Warranty Policy" is useful. "Warranty Policy — Accidental Damage" is what stops the reader having to skim eleven pages, and it is the difference between a citation people click twice and one they click once.
- Do not cite the conversation. If the answer came from something the visitor said three turns ago, that is not a source, it is context. Badging it dilutes the signal.
- Watch for near-duplicate documents. If two versions of the same policy are both in the knowledge base, the badge will name one of them roughly at random, and it may not be the current one. This is a curation problem rather than a citation problem, and it is covered in keeping a knowledge base accurate.
The underlying principle
An assistant that cites its sources is making a smaller promise than one that does not, and keeping it. It is not claiming to know things. It is claiming to have read your documents and to be able to show you where.
That is a promise you can actually keep, which is why it holds up under scrutiny from exactly the people you most want to convince. You can upload a document and watch citations appear on real answers in the dashboard at hiroi.ai.