Conversion Incidents

Pro Feature

Catch conversion drops the moment they happen. The detector watches every active goal and funnel on your site, opens an incident when a metric crosses your baseline, and notifies the right people through the channel they already use.

---

What gets detected

The detector runs every 30 minutes and watches two categories of metric:

• Goal completions (custom_event, url_match, element_click goals you define in Settings → Goals)
• Funnel completion rates (multi-step conversion paths you define in Settings → Funnels)

When the last 60 minutes of completions falls more than 25% below the rolling 7-day hourly average AND the metric has at least 5 completions per hour on average, an incident opens automatically.

Severity is set by the magnitude of the drop:

• Critical when current is less than 40% of baseline
• Warning when current is less than 60% of baseline
• Info for smaller but still notable drops

Incidents auto-resolve when the metric recovers above 90% of baseline. If the same metric drops again within 24 hours, hysteresis prevents a second incident from opening for the same root cause.

---

How AI triage works

Two AI surfaces work together:

Rule-based detector (Phase E, runs every 30 minutes) Computes baselines and deltas in pure SQL. No AI cost. The triage summary is templated from the actual numbers (current value, baseline, delta percent) plus a list of common causes to investigate.

Pattern analyzer (Phase G, runs once daily) A single GPT-5.4-mini call per Pro+ website analyses 24 hours of aggregated telemetry (top paths, referrers, device mix, funnel rates, goal counts) and surfaces up to three behavioural patterns the rule-based detector cannot see. Examples:

• "Users from /pricing return to / before signing up 312 times in 24h, 3x the baseline"
• "Mobile Safari users abandon checkout at the payment-method step at 2x the desktop rate"
• "Tuesday Twitter referral traffic converts 3x the average — consider boosting Tuesday spend"

Patterns appear as pattern_insight incidents in the same dashboard. They route through the same notification channels. Confidence below 0.65 is dropped silently rather than spamming the inbox.

---

Notification channels

When an incident opens (manually or via the auto-detector), the system fans out to four channels in parallel:

Email

Every team owner and admin receives an email by default. The email includes:

• Severity-coloured pill (red for critical, amber for warning, blue for info)
• One-sentence headline with the metric, magnitude, and baseline
• AI triage paragraph
• "Open in dashboard" CTA button
• "You're receiving this because you're an admin of [domain]" footer with a one-click opt-out link

Each user can opt out individually in Profile → Preferences → Conversion Incident Alerts.

Webhook

If you have a webhook configured in Settings → Integrations → Webhooks with the conversion_incident.opened event subscribed (or * for all events), it receives an HMAC-signed POST with the full incident payload. Use this to forward to Slack, Discord, Zapier, n8n, or your own systems.

Mobile push

If you have the Zenovay mobile app installed and registered push tokens, you'll receive a push notification with the metric label, severity, and a deep link to the incident.

In-app inbox

The bell icon in the bottom left of the sidebar opens the inbox. Click the Notifications tab to see incidents alongside other personal notifications (usage warnings at 80%/95%, billing alerts, goal completions). Unread badges sum across both tabs.

---

Setup

1. Define goals or funnels

The detector cannot watch what you have not defined. Open Settings → Goals or Settings → Funnels and add at least one of each kind you care about.

If you have not yet defined any, the Incidents tab on each domain dashboard shows up to three auto-extracted funnel suggestions based on the patterns it sees in your traffic. One click on Track this funnel creates the funnel and adds it to the detector's watch list.

2. Confirm notification preferences

Open Profile → Preferences → Email Notifications and confirm Conversion Incident Alerts is on (it defaults to true for owners and admins). Toggle off if you want to mute these for yourself.

3. Wait for baseline

The detector needs at least 7 days of normal traffic to compute a stable baseline. Until then it skips the metric to avoid false positives. After the first 7 days, drops greater than 25% open incidents automatically.

---

Manual incidents

Sometimes you spot a problem the detector hasn't picked up yet (a deploy you know broke something, customer reports of a checkout error). Click + New incident at the top right of the Incidents tab to log one manually.

The modal has three picker modes:

• Goal: pick from your defined custom goals
• Funnel: pick from your defined funnels (with optional step selection)
• Custom: free-form metric label for ad-hoc incidents

The same notification fan-out fires for manual incidents — your team gets emails, webhooks fire, the inbox shows the entry. This is useful for incident communication ("a fix is being deployed, ETA 15 min") even when the detector wouldn't have caught it.

---

Plan tier behaviour

Free teams are filtered out at the database level — they incur no AI cost and see no incidents in their dashboard.

---

Cost transparency

We expose the cost containment design because it matters:

• Phase E rule-based detector: pure SQL, zero AI cost, runs every 30 minutes per Pro+ website.
• Phase G pattern analyzer: one GPT-5.4-mini call per Pro+ website per 24 hours, hard-capped at max_completion_tokens=600 (approximately $0.005 per call ceiling). Per-pattern dedup prevents the same insight repeating day after day. Sites with fewer than 100 sessions in 24h are skipped (no signal to analyse).
• Routing through Cloudflare AI Gateway with cached/observed token usage so we can see total cost per day in the Gateway dashboard.

Total estimated AI cost at current Pro+ team count: under $30 per month across all customers.

---

Privacy

The pattern analyzer sends only aggregated counts to GPT-5.4-mini:

• Top 8 paths with session counts
• Top 8 referrers with session counts
• Device and country mix breakdowns
• Funnel completion rates
• Goal completion counts

It never sends raw events, raw URLs with query strings, IP addresses, user identifiers, session replay payloads, or any PII. The manifest is bounded to approximately 5-10 kB per call.

---

Related

• Goals — define custom goals the detector watches
• Funnels — define multi-step conversion paths
• Webhooks — receive incident events in your own systems
• Notification preferences — per-user opt-out