Why every chatbot answer should link to a real source

A citation makes the answer checkable

When a chatbot answers a question, you're trusting it to have gotten something right. A citation changes that. Instead of asking you to take the answer on its word, it points you to the actual page the answer came from, so you can look for yourself in a few seconds instead of guessing whether the chatbot got it right.

That's the appeal in one sentence: a linked source turns a claim into something checkable. Without one, you're left weighing how confident the chatbot sounded against how much you're willing to trust a tool you can't see inside.

A citation only works if it actually checks out

Here's the part that's easy to miss: having a link at all isn't the same as having a real one. A citation can point to a page that no longer exists, a page that's close but not quite the right one, or a page that's live and reachable but doesn't actually say what the chatbot claimed. All three look identical at a glance. None of them hold up once you click through.

A chatbot that produces a plausible-looking but wrong URL is doing something specific: reconstructing what a link probably looks like, the same way it can reconstruct what an answer probably sounds like. The link can be wrong in the exact same way the words around it can be wrong, and for the same reason.

A broken citation is worse than no citation

A chatbot that answers with no source at all is at least honest about what it's giving you: an unverified claim, and you know to treat it that way. A chatbot that answers with a source that doesn't hold up is doing something worse. It's borrowing the credibility of a citation, the visual signal that says "this has been checked," without the part that actually makes it checkable. You're more likely to trust the answer because it looks sourced, right up until you click through and find out it isn't.

That's why the presence of a link isn't really the thing worth caring about. What matters is whether the link survives being clicked.

You: Does the annual plan renew automatically, or do we have to re-purchase it each year?

Chatbot: Annual plans renew automatically unless you cancel at least 30 days before the renewal date. You can find this under the Billing section of your account settings, in the "Renewal and cancellation" article.

You: I checked that article and it only talks about monthly plans.

Chatbot: You're right to check, and I don't want to point you somewhere that doesn't actually back this up. I don't have a specific article confirming the annual renewal window, so treat my first answer as unconfirmed until you can verify it with your team directly.

Notice what happened in the second turn. Instead of defending a source that didn't hold up, or quietly swapping in a different-sounding article, the chatbot backed off the claim once it couldn't stand behind where it came from. That's the behavior worth watching for: not just whether an answer comes with a source attached, but whether the chatbot treats "I can't actually back this up" as a real, sayable answer.

What to check instead of just checking for a link

A quick way to test this yourself: ask a question specific enough that only one article could answer it, then actually open whatever the chatbot points you to. Read the specific page or section it named. If the claim is really there, you've confirmed both that the chatbot is grounded and that its citation is trustworthy, not just present. If the source doesn't say what the chatbot claimed, or it's unclear which part of the source supports the claim, that's worth flagging even if the answer itself sounds reasonable.

Try this across a few different questions before trusting a chatbot's citations in general. A chatbot can get the easy ones right, cite them correctly, and still produce a fabricated link on a harder question where it's less sure of itself.

How we handle this

When our agent answers from your documentation, it names or links the specific page the answer came from. If a citation doesn't point to a page we actually retrieved and can confirm, we don't ship it as a working-looking link that might not hold up. We either correct it to the real source or drop the link and keep the plain text, so a broken or fabricated-looking citation can't reach whoever's asking. A plain "I don't know that" doesn't get a citation attached to it either, since there's nothing to back up in the first place.

One exception: the chat box on our homepage runs on this site's own pages, a two-page corpus rather than a real help center, so it doesn't link citations there, since a two-page corpus would mostly just point you back to pages you can already see.