Explainer · Methodology
How we find anomalies
Not a black box. A baseline, a window, and a threshold
ainpulse watches every property's normal pattern, day in, day out — then flags what doesn't fit. Here's the mechanism in plain terms: how we learn what's normal, how we judge what's not, and how we keep the noise out.
Live monitoring
Three properties. Two normal. One needs attention.
The problem
Spotting anomalies isn't spotting drops
Raw analytics data is noisy. Sessions swing day-to-day. Conversion rate fluctuates by 5-10% naturally. Some metrics dip on weekends, lift on Mondays, shift on holidays.
A "20% drop" can be normal. A "5% drop" can be a real regression — if it's outside the property's pattern. Detection isn't about chasing big numbers. It's about distinguishing signal from noise, every day, across every account, faster than a human reviewing dashboards.
What you see in the dashboard
What actually matters
The same data. Two readings.
The job is filtering. The skill is calibration. ainpulse does both, every account, every day.
How we learn
Every property gets its own baseline
Most monitoring tools ship with global thresholds: "alert if sessions drop 25%." That works for nothing.
A property doing 50,000 sessions on a quiet Sunday isn't comparable to one doing 5,000 on a busy Friday. Same number, different meaning. So we don't ship thresholds. We learn each property's normal pattern from its own recent history — separately, for every metric.
28-day rolling window
Illustrative. Each property gets its own. The baseline updates daily.
What the baseline understands
✓It tracks
- Weekday vs weekend rhythm
- Recurring weekly seasonality
- Slow trends (gradual growth or decline)
- How much day-to-day jitter is natural
—It ignores
- Past anomalies (a spike isn't "the new normal")
- Prior campaign spikes still in the baseline window
- External events the property doesn't normally see
- Arbitrary thresholds set by managers or clients
How we judge
Deviation size, made actionable
When actual data drifts outside the baseline zone, ainpulse marks it. But not every miss is created equal. A slight dip below the zone is one thing — a 60% collapse is another.
We translate raw deviation into three tiers, so you don't have to stare at numbers to figure out what needs your attention first.
Reading the deviation
Illustrative. Magnitude of deviation determines tier.
What it means: Outside baseline, modest deviation.
Example: Sessions dropped 18% on a Tuesday vs the property's own typical Tuesday.
Typical action: Worth a look. May resolve on its own, may be early sign of something.
What it means: Substantial deviation, sustained more than a day.
Example: Conversion rate dropped 45% for two days.
Typical action: Investigate today. Real issue likely.
What it means: Major deviation, almost certainly an incident.
Example: Sessions dropped 87% overnight. Tracking visibly gone.
Typical action: Drop what you're doing.
Same logic, three volumes. You see the number; you read the tier; you decide what's worth your attention.
See your own baseline form
Connect one GA4 property in under five minutes — the first checks run tomorrow.
How we keep the noise out
Most alerts you don't want, you don't get
Detection without filtering is just an alarm that screams every Tuesday. The hard part of monitoring isn't catching anything that deviates — it's catching only what's worth catching.
A few mechanisms keep the channel clean so when ainpulse reaches out, it means something.
Day 14: single-day dip
Could be a blip, a redirect glitch, a one-time event.
Suppressed — single-day deviation
Days 14-16: sustained pattern
Pattern says the regression is real.
Alert fired — sustained deviation
Same deviation. Different shape. Different decision.
Sustained pattern
Moderate deviations don't fire on day one — ainpulse waits a day or two to see if the pattern persists, because most blips resolve themselves. Severe failures don't wait: data going to zero, a hard cost spike, or a critical-magnitude drop is flagged on the next daily check.
Volume floor
Tiny absolute changes on small properties don't fire. A property doing 12 sessions doesn't need an alert for an 87% drop.
Alert deduplication
The same alert won't ring twice in quick succession. Once you've been notified, you're not notified again until the situation materially changes.
Quiet by design. When ainpulse pings, it's because the data — not the math — said so.
When you get pinged
An alert with everything you need to act
Every alert carries the same payload: which property, which metric, what changed, how much, when, and how serious. Same context whether it lands in email or Slack. No "something changed" filler — click through, see the data, decide.
Detected: 2026-05-04
Sessions dropped −87% vs baseline (4,200 → 540).
Sessions dropped −87% vs baseline (4,200 → 540).
Different channels. Same content. Either way, you don't have to look for it.
Questions
Common questions
Now you know how it works.
Try it on your data
Connect one property in under five minutes. See your own baseline form. See how the first alert reads when it lands.