Emerging topics and drift: catching new questions before they spike

What is the difference between an emerging topic and drift?

Both describe customer demand changing faster than your knowledge base. They look different in the data, so they need different detection.

• An emerging topic appears quickly. A product launch, a pricing letter, an outage, a new regulation or a change to a form produces questions that weren't there last week. The first few days decide how expensive it becomes, because agents improvise until an answer is written.
• Drift is gradual. The share of contacts about one topic edges up a little every week. Customers start using a new word for an old problem. Re-contact on a topic rises a few points a month. No single day looks unusual, so nothing alerts.

Emerging topics are usually knowledge gaps in waiting: the article doesn't exist yet. Drift more often means existing knowledge is ageing: the article still exists but describes a world that has moved.

Why do monthly dashboards miss emerging topics and drift?

A monthly dashboard answers "what happened last month?" in totals. That framing hides both kinds of change.

For emerging topics, the problem is timing and averaging. A topic that appeared on the 20th has ten days of volume diluted across a full month. By the time the report is read, the topic has been costing money for several weeks and agents have settled into their own answers, which may not agree.

For drift, the problem is the comparison. Month against previous month shows a small change, which looks like noise. Nobody lines up six months to see that every one of those small changes went the same way.

Category design adds a third blind spot. A genuinely new question rarely has its own category, so it lands in "Other" or in the nearest existing category, where it raises volume slightly and looks unremarkable. You only see it by looking at what customers said, not just at the category totals.

What makes a fair baseline?

Every alert is a comparison, so the baseline decides whether an alert means anything. Contact centre demand has strong weekly patterns: Mondays are busy, weekends are quiet, and some topics cluster around billing dates. Comparing Monday with Sunday reports a spike every week.

A fair baseline compares like with like:

• Same weekday last week for daily checks. Tuesday against last Tuesday absorbs the weekly cycle.
• The previous full week or month for weekly and monthly checks.
• A normal range from several periods, not a single one. If last Tuesday was itself unusual, comparing against it alone produces a false spike or hides a real one. Checking today against several previous Tuesdays shows whether it is outside the usual spread.
• A holiday calendar for each region you serve. Public holidays and the days around them distort volume. Holiday days should be flagged, kept out of the normal range, and explained in reporting instead of alerted on.

Time zones matter too. A team serving customers in Perth and Sydney needs its days cut at local midnight, or a busy evening lands in the wrong day.

Which alerts catch sudden change?

A small set of alerts covers most sudden changes in a contact centre. These are the types Knowledge Quest raises, with its default thresholds, which can be adjusted for each knowledge area:

The last two are the ones most dashboards lack, and they matter most to a knowledge team. A volume spike tells you more people called. A knowledge gap surge tells you more people asked something your articles couldn't answer.

How do you catch drift that never trips an alert?

Thresholds are tuned to ignore small movements, which is exactly why drift slips past them. Lowering the thresholds just produces noise. Look across several periods instead:

1. Read trends over eight weeks or more. A topic up 4% in each of six weeks has grown by more than a quarter. Plot share of total volume as well as raw counts, so overall growth doesn't hide a shift in mix.
2. Watch the words, not just the categories. Track terms customers use that are new or rising against the previous equal window. A new product nickname, a competitor's offer or a regulator's phrase often appears in conversations before it has a category.
3. Compare customer vocabulary with your articles. Words customers use that your articles never mention are a slow-moving findability gap. Each one makes the right article a little harder to reach.
4. Track the unmapped share. A creeping rise in contacts that fit no category is drift in the taxonomy itself. Our guide to building a taxonomy from conversations covers how to repair it.
5. Track outcomes by topic. Resolution and re-contact drifting the wrong way on one topic, while volume stays flat, usually means the article for that topic has aged. The same measures tell you whether a fix worked, as set out in measuring knowledge article impact.

A short weekly review that walks through these five views takes less time than investigating one false alarm, and it catches the changes that alerts are designed to ignore.

What should happen when a new topic appears?

An emerging topic needs a decision, not just a notification. The quickest way to choose one is to check two things: how much of the topic your articles already answer, and how its conversations are ending.

• No coverage: create an article. Agents are improvising.
• Partial coverage: review what exists. Some intents are answered and some aren't.
• Covered, with poor outcomes: improve the article. Low resolution or high re-contact on a covered topic means the answer is wrong, incomplete or hard to follow.
• Covered, with good outcomes: monitor it. The knowledge base is coping.

Prioritise topics where coverage is low, resolution is weak or re-contact is high. Draft the answer from the conversations that raised the topic, because they contain the exact questions and follow-ups customers have. If the topic is temporary, such as an outage, write a short-lived article with a review date and retire it when the topic fades. The knowledge gap analysis guide covers how to size and close each gap.

How do you stop alerts becoming noise?

An alert stream people ignore is worse than none. A few rules keep it credible:

• Set minimum volumes. Percentages on small numbers are unstable. Six contacts against a baseline of one is a 500% rise and may mean nothing. Show the count and flag small samples.
• Hold back anything within its normal range. If today sits inside the spread of recent comparable days, don't alert, however far it is from one baseline day.
• Downgrade on holidays. Keep holiday-affected changes visible at lower severity, with the holiday named.
• Cool down after acknowledgement. Once someone has acknowledged an alert, don't raise the same one again for the same scope straight away.
• Say what held steady. A summary that states which measures stayed within normal range stops readers assuming the worst about everything not mentioned.

How does Knowledge Quest watch for emerging topics and drift?

Knowledge Quest reads every conversation from your contact centre platform, such as Genesys Cloud or Amazon Connect, extracts each customer's intents and classifies them against your taxonomy. It compares each period with a weekday-matched baseline, checks metrics against their normal range over previous comparable periods, and applies a holiday calendar for each knowledge area. It raises the alerts above, and reports emerging topics with the recommended action, priority and a recent trend for each. Daily, weekly and monthly briefs explain what changed and what held steady.

For drift, Kai, the built-in knowledge analyst, answers questions such as "is this topic trending up over the last eight weeks?", "what terms are new or rising?" and "which words do customers use that our articles don't?", with the sources behind each answer. None of it replaces your contact centre platform or knowledge base. It makes sure the knowledge base hears about change while it is still small.