Half of this vocabulary was borrowed from teams with a ticket queue

The half worth watching, and the half hardest to reduce to one number.

• deflection rateshown in the productThe Deflection tile is the share of assistant messages that were not refusals. Support teams usually mean something wider by the word.
• fallback rateshown in the productThe refusal share is the other half of the Deflection tile. What triggers a fallback, what it costs, and why a low one is not automatically good.
• coverage gapsshown in the productEvery refused question is a documented hole in your knowledge base, written by a real visitor. How the list is built and how to work through it.
• retrieval scoreshown in the productThe similarity between a question and the nearest passage decides whether an answer happens at all. What the figure is, and what it is not.
• answer latencyshown in the productThe latency tile measures machine time to a finished answer, not how long a visitor waited to see anything. Why those differ, and which one matters.
• satisfaction on automated repliesrecorded, not on the insights pageA visitor can rate a conversation up or down. The counts exist, no percentage is computed, and here is why that restraint is deliberate.
• resolution ratenot measured hereResolution requires somebody to judge that a problem ended. A widget sees a conversation stop, which is a different thing entirely.
• abandonmentnot measured hereNobody records a closed panel or a visitor who never came back. The outcome that matters most here leaves no trace at all.
• handover qualitynot measured hereAn enquiry arrives as a name, an address and a message. Nothing grades it, and the conversation behind it is not in the email.

Easy to count, easy to mistake for progress.

• conversation volumeshown in the productHow many people opened the widget and said something. A useful denominator, a poor headline, and a number that moves with your marketing.
• containment ratenot measured hereContainment counts conversations that ended without a person. A widget cannot see whether one did, so no containment figure exists here.
• first response timenot measured hereThis metric measures how long somebody waited for a human to pick up. There is no queue in a widget, and no first response time is computed.
• escalation ratenot measured hereHandover enquiries are stored and listed. No rate is computed against them, and dividing by conversations by hand needs three caveats.
• repeat question raterecorded, not on the insights pageWhen somebody rephrases a question they already asked, the first reply failed. Every message is stored and no repeat is counted.

The only figures here with no interpretation in them.

• cache rateshown in the productSupport questions repeat. When a new question means the same as a recent one, the previous answer is returned. What that share includes and excludes.
• cost per answershown in the productThe page shows what was spent and how many messages were answered. Dividing them is the useful number, with three caveats about the denominator.
• reply allowance burnshown in the productThe plan meters replies. Refusals and cached answers do not spend one, a greeting does, and the ceiling is not a friendly notice.

Words this field uses loosely, defined so a conversation can be had.

• groundingshown in the productGrounding 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 a support contextnot measured hereA confident wrong answer about a refund window is a commitment somebody will hold you to. No product detects these, including this one.
• intentnot measured hereIntent means what the visitor is actually after, which is rarely what they typed. Nothing here labels or counts it, and here is why.
• confidence thresholdshown in the productOne number decides whether a question gets an answer or the refusal message. Three settings, what each does, and what each costs.
• knowledge freshnessrecorded, not on the insights pageSyncing and the one day answer cache bound how stale a reply can get. No staleness score is computed, and a synced document can still be wrong.
• time to first answer after publishingnot measured herePublishing a page does not make it answerable. What has to happen in between, and why the wait is longer than most people expect.
• source coveragenot measured hereCoverage is a property of your documents rather than of the assistant. Nothing computes it, and here is the nearest evidence.

A glossary published by a vendor tends to define the terms that vendor happens to measure and go quiet on the rest, which leaves a reader assuming a dashboard that does not exist. 10 of the pages here are about things this product does not compute at all. They say so in the first box on the page, and then set out what would be needed to measure the thing properly and what to use in the meantime.

If you want the argument rather than the definitions, the writing on what is worth measuring and the containment rate trap is the place for it.