What is Trends?

Trends is a beta page that answers a narrower question than the rest of Analytics: what’s moving right now, and is it unusual? Instead of a dashboard you read for context, Trends runs statistical tests against a chosen time window and hands back a short list of signals — each one backed by figures, a channel breakdown, and sample conversations you can check for yourself.

It’s currently in beta, sitting behind a feature flag set per workspace. Your workspace may not have it yet.

The detectors don’t run when you open the page: a scheduler counts and tests every 5 minutes, and the page reads the latest result (cached for a minute), so what you see may be up to a few minutes old. But Trends is not only a page: a spike that clears a stricter bar than the page uses also raises the Trend Spike Detected automation trigger, so a rule can react to it without anyone watching. Trends never delivers a message itself; it raises the trigger, and the automation you build decides what happens — a Slack DM, an email, a webhook, an assignment. See Build a Trend Spike automation.

What Trends actually tests

Trends runs two detectors:

• Volume — is the tenant’s total conversation count in this window unusual, compared against a seasonal profile built from the tenant’s own history by day-of-week and hour?
• Label — given the volume actually observed in the window, is a particular label or tag taking an unusual share of it?

The label test deliberately conditions on the volume actually seen, not a seasonally predicted one. If total volume doubles for an ordinary reason and every label keeps its usual share, that’s a volume story, not a labels story — Trends reports it once, on the Volume card, rather than flagging every label at once.

The two tests count different things, so read each number in its own unit. A volume signal counts conversations. A label signal counts contacts, de-duplicated inside each 15-minute arrival bucket and then summed — so one person who comes back in a later bucket is counted again. The window total the share test conditions on is always a count of conversations.

Reading the page

• Time window selector — choose the window you want tested. Trends re-runs both detectors against whatever window you pick.
• Volume card — shows the tenant’s total for the window and whether it reads as statistically unusual against the seasonal baseline.
• Signal cards — one per detected anomaly (label or otherwise), each carrying:
• a severity indicator,
• a headline summarizing what’s unusual,
• the underlying figures (counts and shares),
• a channel breakdown,
• and a handful of sample conversations so you can see, not just be told, what’s driving the signal.

“No signals” vs. “not enough data”

Trends never shows an empty signal list as if that meant nothing was happening. Instead, the page is explicit about which of a few states you’re in:

• No signals — the tests ran and found nothing unusual in this window.
• Not enough data — the window doesn’t hold enough traffic for a share test on a given label, and the page says so rather than silently omitting it.
• Signals left out — if some signals were excluded, the page states how many.
• Backfill in progress — if historical data is still loading, the page tells you the backfill hasn’t finished rather than presenting a partial baseline as complete.

Each of these is meant to stand in contrast to a quiet, empty page — Trends always tells you why the list looks the way it does.

Alerting vs. the page

The page and the alert read the same detections; what separates them is how strict the bar is. Anything statistically interesting is worth showing on a page, because a page costs nobody anything. Interrupting somebody is different, so an episode only alerts if it also clears a much lower p-value, a minimum effect size, a minimum excess in contacts, and elevation across at least two separate 15-minute arrival buckets — that last one so a sliding window re-reading the same conversations can’t be mistaken for a second piece of evidence.

An episode alerts once, updates only if it materially worsens, and goes quiet again when the label does. There’s a cooldown between messages about the same episode.

Those thresholds were calibrated against 62 days of production history across four tenants, tuned to a budget of a few alerts per tenant per week. They’re deliberately where they are: one step stricter measurably cost real detections and slowed down time-to-detect.

So: use the page when you want to go look at what’s unusual, and build a Trend Spike Detected rule when you want the unusual to come find you.

For the rest of Atender’s reporting — historical dashboards, SLA and CSAT views, agent performance, and more — see What is Analytics?. For a real-time pulse of the last 24 hours, see The Dashboard.