An answer you cannot check is a rumour

There is a version of this product that answers slightly more fluently, looks cleaner, and is much harder to trust. We keep choosing against it, and the reasoning is worth writing down.

The failure that citation actually prevents

A customer emails: your bot told my client we ship to Canada. You do not ship to Canada.

Without a citation you now have an investigation. Did someone write it on a page? Is it in an old PDF? Did the model invent it? You cannot reproduce it, so you cannot fix it, so you apologise and hope.

With a citation you have a link. Either it points at a 2023 shipping page nobody deleted — in which case you delete it, and the problem is permanently gone — or it points at nothing relevant, in which case you know the retrieval step failed and that is a different, also fixable, problem.

Same wrong answer. One is a mystery, the other is a ticket. That is the entire argument.

Why a model has to be told what is true about you

A language model has no knowledge of your business. None. It has read an enormous amount of text about businesses in general, which is why it can produce a fluent, plausible, completely invented refund policy without any sense that it is doing something wrong.

So the answer has to be built from your content: find the relevant passages first, then ask the model to answer from them. Once you have done that, you already know which passage the answer came from — showing it costs almost nothing. Which raises the real question.

Why would anyone not show it?

Three reasons, and none of them is good for the person asking.

• It looks tidier. A bare paragraph is cleaner than a paragraph with a link under it. This is a real design cost and we pay it deliberately.
• It reveals when the source is thin. An answer citing a page that barely supports it is visibly weak. That is information the reader is entitled to.
• It invites correction. Citations generate emails saying “this page is wrong”. Vendors experience that as support load. It is the most valuable feedback in the system.

What citation does not fix

The part usually left out of posts like this one, because it undercuts the pitch.

• A confidently wrong answer from a correct source. If the assistant misreads a passage, the citation points at a page that genuinely exists and genuinely says something adjacent. Now the wrong answer looks better sourced than it is.
• An out-of-date page. Cited faithfully, wrong anyway. Citation surfaces this eventually, but only when someone notices.
• The question it should have refused. Nothing about citing sources tells an assistant that a question is outside what it should be answering at all.
• People not clicking. Most will not. The citation still works, because its main function is that the answer had to come from somewhere — but let us not pretend everyone checks.

What we do about the first one

The dangerous case is a wrong answer wearing a real citation. Two things help, and neither is a solution:

• Say when it does not know. An assistant that will admit a gap produces fewer confident misreadings than one that must always produce something. The demo on the landing page includes a scenario where the assistant fails — it is the fourth one, and it is there on purpose.
• Show the passage, not just the page. A reader can see in two seconds whether it supports the claim. A page link takes two minutes, so nobody does it.

The version we did not build

An assistant that speaks with total confidence and cites nothing tests better in a five-minute demo. It feels more capable. It is more impressive to a buyer who is not going to read the transcripts.

It is also unimprovable, because every failure arrives as an anecdote instead of a reproduction. We would rather ship the version that occasionally says “I do not have that” and always says where it got the rest.

The mechanics of the retrieval step are on how it works, and what the models receive is in the AI transparency statement.