Question 1

What does Trends actually detect?

Accepted Answer

Two things, kept apart. Total volume: whether the number of conversations arriving is unusual for this day and hour, against your own seasonal profile. And composition: whether a topic is taking an unusual share of the traffic you're getting right now. The second catches an outage in the first twenty minutes — a login problem needn't move your total volume at all to be obvious in the mix.

Question 2

How is this different from Analytics?

Accepted Answer

Analytics is a dashboard you go and read: what happened, how you're doing, over whatever range you choose. Trends has already done the reading and answers what is different right now — the last 2 hours to 7 days, recomputed every five minutes. And it only puts something on the page when the counts clear a bar.

Question 3

How quickly does something show up?

Accepted Answer

Counts are bucketed at 15 minutes and the job runs every five, so a spike surfaces roughly 15 to 20 minutes after the conversations start arriving. The page says when it was last worked out.

Question 4

Won't a busy morning set off an alarm for every topic at once?

Accepted Answer

No — avoiding that is the central design decision. Each topic is tested against the volume you actually received, not a prediction of it. If your traffic doubles and every topic keeps its share, the volume detector reports one event and every topic stays silent.

Question 5

Does one angry customer opening ten tickets look like a trend?

Accepted Answer

No. Trends counts distinct people, not conversations — ten conversations from one contact count once. It's also why the numbers are smaller than a ticket count and shouldn't be compared to one. The card names the unit every time.

Question 6

Where do the numbers on a signal come from?

Accepted Answer

Straight from the counts, with no model involved: how many customers, what's usual for the topic, the resulting multiple, its share of the window, and roughly when it started. Underneath: the biggest 15-minute bucket, the conversation total, a per-channel split, and the conversations themselves.

Question 7

What stops a long outage from poisoning what counts as normal?

Accepted Answer

Windows already flagged as incidents are taken out of the 28-day baseline, with a guard band around them — otherwise a week-long problem teaches the detector that the problem is normal. Removal is capped at about a fifth of the sample so the baseline can't walk downward indefinitely. Over 62 days of production traffic, the cap cost nothing in recall.

Question 8

What happens when there isn't enough traffic to say anything?

Accepted Answer

It says so. Below 30 conversations in a window a share test can't separate a real change from noise, so per-topic detection switches off and the page reports total volume only. The same applies while it's still reading your history, and to anything left off the page — omitted signals are counted, never dropped in silence.

Question 9

Does it depend on us tagging conversations properly?

Accepted Answer

It detects on the tags you already have, and states how much of each window carries one, so you know what share of your traffic it could see. Total-volume detection needs no tags at all.

Question 10

Who in our team can see Trends?

Accepted Answer

Owners, team leads and analytics roles — not agents, not readers. Trends shows customer text from across the whole workspace, which an agent doesn't normally see, so it has its own permission rather than riding on an existing one.

Be the first to know what's breaking

Every number, and where it came from

It knows the difference between a busy morning and a bad one

Yesterday's incident doesn't become tomorrow's normal

It tells you what it cannot see

Conversations

Analytics

Incidents

Monitor

What changed, and how sure are we.

See Atender in action