What is this chat about, at a glance
When an agent glances at a conversation they did not start, the first question is always the same: what is this one about? A transferred chat, a tab they left twenty minutes ago, a backlog of waiting visitors. Before anything else can be decided, the subject has to be established.
Today the only way to answer that is to read the transcript, and reading is expensive at exactly the moment when attention is scarcest. The topic badge collapses that scan into a glance. It watches the messages, matches them against curated keyword sets for nine common support subjects, ranks them by how many messages touched each, and shows the leading one with a confidence label and the keywords that led it there.
It fills a genuine gap on the rail. Other cards measure dimensions of a conversation, track items within it, rank tabs against each other, or hold reference context. None of them own the categorical question of subject. A buying-intent score is not a topic, and a lifecycle phase tells you where you are in the arc rather than what the arc is about.
Nine topics, and both sides of the conversation
The nine subjects are Outage, Cancellation, Billing, Account, Shipping, Technical, Sales, How-to and Feedback. They are ordered roughly by specificity and urgency, and that ordering does real work: it breaks ties, so when two topics score equally the more specific or more urgent one is the one you see. A chat that reads equally as Outage and Technical surfaces as Outage, which is the correct way round for something you might need to act on.
Unusually for the rail, both the visitor's messages and the agent's are scanned. That is a deliberate departure from cards such as the mood meter or the self-stated facts list, which read the visitor only. A subject emerges from a whole conversation rather than one side of it. A visitor who opens with “this isn't working” has named no topic at all; your reply asking about their last invoice is what makes the chat a billing chat. Ignoring the agent's half would systematically under-classify precisely the conversations that started vaguely.
Each topic also keeps a short trail of the keywords that matched, capped so a noisy chat stays readable. The trail is the feature that makes the badge auditable: when a label looks wrong, the words that produced it are right there, and you can usually see the misfire immediately.
One message is one vote
The scoring rule is the most important detail in the card, and it is a single sentence: a topic scores one point per message that matches it, no matter how many of its keywords appear in that message.
The alternative would be to count every keyword hit, and it fails in a way that is easy to picture. A visitor pastes a long, frustrated paragraph containing the words invoice, charge, refund, receipt and card. Under hit-counting, that one message makes the chat overwhelmingly about billing. Meanwhile a genuine outage, mentioned once per message across fifteen turns because it is the actual problem, scores lower. Volume in one message would beat recurrence across the conversation.
Presence-based scoring inverts that. It measures how much of the conversation touched a subject rather than how densely one message did, which is much closer to what a person means when they say what a chat was about. A single keyword-stuffed message can contribute at most one point to each topic it mentions.
Clear, Likely and Mixed
Alongside the leading topic sits a confidence label with three values, and reading it is what separates useful triage from a misleading badge.
Clear means either only one topic matched at all, or the leader has at least two matching messages and at least twice the runner-up. Those are the chats that are genuinely about one thing.
Likely means the leader is ahead, but not by that margin. The badge is probably right and you should not lean on it for anything consequential without a glance at the trail.
Mixed means the top two are tied. This is not a failure. Plenty of real conversations are legitimately about two things, and a visitor who wants to cancel because they were double-charged is a cancellation chat and a billing chat at once. Mixed is the badge telling you the truth about a conversation that does not have a single subject, which is more useful than a confident guess at one of them.
Why it waits for a second signal
Until the accumulated topic score reaches two, the card reads as warming up rather than naming a subject.
A single keyword is not a topic. The word account appears in an enormous range of conversations that are not about accounts, and labelling a chat from one match would produce a badge that flickers between subjects for the first few messages and then settles. Agents learn very quickly to ignore a signal that changes its mind three times, and once ignored it may as well not be there.
Waiting for a second matching message is a low bar that removes most of that flicker. It is the same restraint the phase gauge and the momentum gauge apply: report nothing until there is enough to report, on the grounds that silence is more honest than noise.
Where keyword classifiers fail
A lightweight keyword classifier has known blind spots, and the card is documented as having them rather than pretending otherwise.
It does not understand negation. “This is not a billing problem” contains the billing vocabulary and will score as billing. It does not detect sarcasm, so a visitor being caustic about how wonderful the outage has been reads as ordinary matching text. And the lexicons are English-only, which means a conversation in another language will sit on warming up indefinitely rather than being classified badly.
These are acceptable limits for a triage gauge and unacceptable ones for a reporting system, which is the distinction to hold on to. The badge is there to answer “what is this about” in the second before you read a chat. If you find yourself citing topic distributions in a monthly report, you have moved past what the classifier was built to support, and the keyword trail is the reason you can tell: spot-check a handful of chats and the negation cases show up quickly.
Conversation tags and the topic badge look similar and are not interchangeable. Tags are applied by a person, mean whatever your team has agreed they mean, and persist as a deliberate record. The topic badge is derived automatically, means only what the lexicon matched, and is a live reading of the conversation in front of you.
The productive relationship between them is that the badge is a prompt and the tag is the record. A badge reading Cancellation on a chat you were about to close as a general enquiry is a useful nudge to tag it properly. Treating the automatic label as a substitute for tagging, though, means your reporting inherits every negation failure and every English-only gap, and it will do so silently.
Like its neighbours, the card keeps state in memory per visitor and drops it when a chat is transferred, accepted, rejected or ended, so the next agent classifies from their own transcript rather than inheriting a stale subject. It adds a topic line to the wrap-up recap when you copy or save it, and deliberately does not touch the readiness check, because every topic is a perfectly valid subject on which to close a conversation.
How MyLiveChat fits
The badge sits at the top of the right-hand rail in the web agent console, with its own chord in the numbered series covered by the keyboard shortcuts guide. It runs entirely in the console and makes no server calls.
If you want subject data you can actually report on, pair the badge with deliberate tagging: tagging and categorisation covers how to build a tag set small enough that agents use it consistently. For what to do with the resulting numbers, the metrics guide is the better starting point than any single gauge.
And when a topic label surprises you, open the card and read the keyword trail before you either trust it or dismiss it. The trail exists precisely so that a classifier without judgement can still be checked by someone who has some.