What a citation has to point at to be worth anything

Imagine two replies to a question about your returns window. Both say thirty days from delivery, both are phrased pleasantly, both are correct today. One of them carries a marker pointing at the paragraph on your returns page that says so. The other does not.

From the visitor's side, in that moment, there is little to choose between them. From the operator's side they are not comparable products. One is a claim; the other is a claim with an audit trail. Everything below follows from that distinction, and none of it is visible on the day everything works.

The distinction becomes stark when the returns window changes to fourteen days. The grounded system either updates when the source updates or points at a passage that visibly says fourteen, so the discrepancy is findable. The ungrounded system continues saying thirty, plausibly and confidently, and there is no artefact anywhere that reveals it. You find out from a customer.

The operational value of a citation is that verification becomes a one minute task instead of an investigation. Open the cited passage, read it, compare. Anybody can do it, no expertise in the tool is required, and it can be done on a phone in a queue.

That cost difference determines whether checking happens at all. Teams do not skip answer review because they do not care about correctness; they skip it because each check costs several minutes of finding the source, and several minutes times a meaningful sample is an afternoon nobody has.

It also changes who can check. A citation lets you hand an answer to the person who owns that subject and ask a narrow question: does this passage say what this answer says. That is a question a busy specialist will actually answer in a message, which is not true of please review our chatbot's replies.

When a grounded answer is wrong, the citation tells you what to edit. Not roughly, exactly: this passage produced this claim. The fix is a specific edit to a specific paragraph, and the same fix takes effect for every future question that lands on it.

Without that, correcting an answer is guesswork. You know the assistant said something wrong. You do not know which of your documents it came from, and it may have come from several at once, or from a stale page you forgot was indexed. The usual response is to add a new page saying the right thing, which leaves the wrong page in place, competing for retrieval, and now the system has two sources that disagree.

This compounds. A year of correcting an ungrounded assistant by adding clarifications produces a body of material with internal contradictions and no map of which page produced which answer. A year of correcting a grounded one produces the material you should have had in the first place.

The failure mode of an ungrounded assistant is a confident, well phrased, wrong answer that nobody notices. It produces no error, no refusal, no complaint. The conversation ends normally. From every reporting angle it looks like a success, and it will keep looking like one.

Grounding changes the shape of the failure. When there is no adequate source, a properly designed system can say so, rather than composing something from the nearest available material. Askably applies a match threshold below which it does not reach the model at all and returns a fallback message the customer has written, which converts an invisible wrong answer into a visible unanswered question.

Visible failures are cheap. They appear in your unmatched questions list, they tell you what to write next, and a visitor who is told plainly that something is not covered is mildly disappointed rather than misinformed. The whole argument for grounding is that it trades a small number of visible failures for a large number of invisible ones, which is a trade worth making every time.

It has to point at a passage, not a document. A citation to a fifty page handbook is not checkable, because checking it means reading the handbook, which returns you to the cost problem the citation was supposed to solve. Passage level is the minimum useful granularity.

It has to be the passage the claim actually came from, not a plausible one found afterwards. A citation retrofitted by searching for something that supports the answer is decoration, and worse than no citation because it creates confidence without providing verification. The distinction is invisible from the outside, so it is worth understanding how your system produces them.

It has to be openable. If the reader cannot get to the source text, the citation is a claim about a claim. This is where citations to internal documents need thought: a numbered marker is fine for a visitor if the underlying passage is available to the team, and misleading if nobody can reach it at all.

And it has to identify a version. Sources change. A citation that points at a page as it is today, when the answer was generated last month against a different version, will send a checker looking for text that is no longer there, which produces a false failure and wastes exactly the time citations were supposed to save.

A citation to the homepage, or to a page that merely contains the topic. This is common and it is theatre: it looks like grounding, it satisfies a glance, and it supports nothing. Anybody who tries to verify it once learns not to bother, which is worse than never having offered.

A citation attached to an answer that in fact came from the model's general knowledge, with the source found afterwards. This is the most damaging variety, because the citation is doing the opposite of its job: signalling verification while concealing that the claim was not derived from the material.

A citation nobody ever opens. This one is on the operator rather than the system. Citations pay for themselves only through use. If nobody on the team has clicked one in a month, the feature is decorative in practice regardless of how well it is implemented, and the answer quality is going unchecked exactly as if there were no citations at all.

It does not make the source true. An answer faithfully grounded in a page that has been wrong since last spring is a faithful reproduction of an error, and citations will not tell you the source is stale. They tell you where the error lives, which is a different and much smaller service, though still the one you need to fix it.

It does not prevent incompleteness. An answer can cite accurately and omit the exception that sits three paragraphs further down. This is the most common quality failure in grounded systems and the one citations are least helpful against, because the citation checks out.

It does not remove the need for escalation rules. Grounding is about where the answer came from, not about whether the question should have been answered at all. A perfectly cited answer to a question that should have gone to a person is still the wrong outcome, and no amount of source transparency changes that.