Anomaly Detector

Scan your product data for things that don't look right — sudden spikes, unexpected drops, metrics that broke from their normal pattern.

When to Use

• "Did anything unusual happen this week?"
• "Alert me if conversion drops below its normal range"
• "Scan all our key metrics for anomalies"
• "Why did page views spike on Tuesday?"
• "Is the error rate abnormally high right now?"
• "Check if any funnel step changed unexpectedly"

Mental Model

An anomaly is a data point that deviates significantly from the expected pattern. Three types:

1. Spike: Sudden increase (errors, traffic, rage clicks)
2. Drop: Sudden decrease (conversion, engagement, revenue)
3. Break: The pattern itself changed — e.g., a metric that was cyclical suddenly went flat, or a metric that grew steadily now oscillates

Use free fullstory:discover_groups and fullstory:get_opportunity_stats for live signals. Discovery accepts one scope at a time (page_id, domain, app_name, segment_id, or funnel_id); use a segment for device/user filters and the live schema for time ranges. Use computed metrics for exact custom windows.

Workflow

Step 1: Choose what to scan

If the user has specific metrics in mind, scan those. If they say "scan everything," pick a default set:

• Page views (traffic health)
• Errors (console errors, network failures)
• Conversion (key funnel completion)
• Frustrations (rage clicks, dead clicks from discover_groups)

Ask: "I'll scan page views, errors, conversion, and frustrations over the last 14 days. That OK?"

Step 2: Build trend metrics

For each metric, build a trend over a wide time window (14-30 days):

fullstory:build_metric(query="page views", output_type="trend")
fullstory:compute_metric(metric_id, time_range="last_30_days")

Step 3: Detect anomalies

For each trend, look for:

Spikes/drops (>50% change from previous period's average):

• "Page views: 12,340 daily average, but Tuesday hit 31,200 (+153%) — possible anomaly unless there was a campaign or launch."

Breaks (pattern change):

• "Checkout conversion was stable at 20-22% for 3 weeks, then dropped to 14% on July 15 and hasn't recovered."

Zero events (metric that had data, now doesn't):

• "Rage clicks on /settings dropped to zero on July 20. Was a fix deployed? Or did tracking break?"

Step 4: Investigate anomalies

For each anomaly flagged:

1. Check if it correlates with a known event (deploy, campaign, holiday). Use the deploy or annotation source available to the user; discover_groups finds frustration signals, not annotations.
2. Narrow the time window around the anomaly point
3. Build a top_n version of the metric grouped by dimension (page, device, browser) to see if the anomaly is concentrated or widespread
4. Pull sessions from the anomaly period using get_sessions(metric_id) to see what users experienced

Step 5: Classify

Step 6: Report

## Anomaly Scan — Last 14 Days

### Found 2 anomalies

1. 🔴 Checkout conversion dropped 14% → 8% on July 15
   - 8 days of sustained lower conversion — not a one-day glitch
   - Co-occurs with v2.3.0 deploy (per annotation)
   - No error spike, no new frustration signals
   → Recommend: run funnel-doctor on checkout flow

2. 🟡 Page views spiked +153% on Tuesday Aug 3
   - Single-day spike, returned to normal Wednesday
   - No error or frustration change
   → Likely external traffic event — not a concern

### 8 other metrics scanned — all normal ✅

Guidelines

• "Anomaly" doesn't mean "problem." A traffic spike from a successful launch is an anomaly, but it's good news. Classify, don't just flag.
• One-day blips are usually noise. Focus on sustained changes (3+ consecutive data points outside the normal range).
• Always check for an annotation or deploy that explains the anomaly. "Conversion dropped on July 15 — that's the same day as the checkout deploy."
• Weekday/weekend patterns are normal. Don't flag "Sunday traffic is lower" as an anomaly.
• If you don't have enough historical data (<7 days of trend), say so: "I need at least 14 days of data for reliable anomaly detection. Right now I can only spot very large swings."