Metrics and terms
What they meant, which nothing here labels
Intent is the most useful idea in support that this product deliberately does not implement. There is no classifier, no set of labels, and no report counting how many people asked about billing. Understanding what intent is remains worth your time, because the work it describes is work you do by reading.
not measured here
This product does not classify or count intent. No labels are assigned to questions, no per intent volumes exist, and nothing in the interface reports what visitors were trying to do. The "Unanswered questions" list groups rows by the question text, so the same intent expressed three ways produces three rows.
What it means
An intent is the goal behind a question, as distinct from the words used to express it. Someone typing where is my order is asking about a delivery date, and someone typing has it shipped yet is asking the same thing in a way that shares almost no vocabulary. Older systems were built around classifying an utterance into a fixed list of intents and then following a script per intent, which made them predictable and made them brittle, because a question outside the list fell into a catch all and got nothing. A retrieval based assistant sidesteps the classification entirely: it matches the meaning of the question against the meaning of your material, which handles unlisted questions gracefully and gives up the neat reporting that a fixed list provides.
How it is actually calculated
How it is done where it is done
Define a list of intents, label a body of real questions, train or configure a classifier, and report volumes per intent. The reporting is genuinely useful and the maintenance is the reason people abandon it.
Every list drifts. New products, new policies, and new phrasings arrive continuously, and a list that is not maintained quietly funnels an increasing share of traffic into its catch all while the dashboard continues to look confident.
What happens here instead
The question is matched against passages by meaning and answered from whatever wins. No label is assigned, nothing is counted per category, and nothing appears in the interface describing what visitors were trying to do.
The unanswered questions list is grouped by the exact question text rather than by intent, which is why the same underlying need appears there several times in different wordings. That is a deliberate limitation and it is where you notice its absence.
The grouping you do by hand
Reading a month of unanswered questions and sorting them into groups is intent classification done by a person, and it usually collapses a long list into a handful of real needs.
Do it by hand for the first several weeks. The groups that emerge from your own traffic are rarely the ones you would have written down in advance, and knowing them is what makes the next batch quick to sort.
How the number gets moved without anything improving
How an intent report flatters a system that is not working
Keep the list short. Fewer intents means higher confidence per intent and a tidier chart, and everything that does not fit lands in the catch all where nobody looks. The report improves as coverage narrows.
Report volumes without reporting the catch all. This is the standard omission, and it is the one number in an intent report worth reading first, because it is the share of your traffic the system has no view of.
The accidental version is drift. A list built once and never revisited looks stable for exactly as long as your business does, and the moment a new product ships the report keeps its confident shape while describing less and less of what is actually being asked.
What to look at instead, or alongside
- The unanswered questions grouped by meaning by hand, which is intent classification aimed at the questions that failed.
- Reading a sample of conversations, where the intent is obvious to a person in about two seconds per conversation.
- The phrasings visitors actually use, added to your material, since matching handles them without any label being assigned.
- Your own site search terms, which cover the same ground for people who never opened the widget.
Questions
- Can I see a breakdown of what people ask about?
- Not in the interface. The nearest available thing is the unanswered list, which shows what failed rather than what was asked, and reading conversations, which shows everything at the cost of your afternoon. The failed list is where the actionable half lives.
- Would intent labels make the answers better?
- Not directly. Matching by meaning already handles the questions a label would have routed, including the ones nobody thought to list. What labels would buy is reporting, which is a real want, and it is a reporting feature rather than an answering one.
- How do I group the unanswered list without doing it by hand every week?
- Do it by hand at first, because the first few weeks teach you what the groups are and there are fewer than you expect. Once the groups are stable, sorting new rows into them takes minutes. The reading is where the insight comes from, so do not automate that part early.
Keep reading
- Grounding: answers bound to passages the model was shownGrounding is a property of how an answer is produced, not a percentage. Here it is enforced by construction and displayed as numbered citations.
- Hallucination in support: what it costs and how to catch itA confident wrong answer about a refund window is a commitment somebody will hold you to. No product detects these, including this one.
- Confidence threshold: the dial that decides refusalsOne number decides whether a question gets an answer or the refusal message. Three settings, what each does, and what each costs.
- Every metric and termWhat each one means, and which of them this product actually computes.
Try it on your own material
Upload a document or point it at your site, paste one line of HTML, then ask it something only your business could answer.