web-analytics

Web Analytics

GA4 extraction, normalized web metrics, seasonal anomaly detection, navigation path analysis, and predictive audience scoring — the behavioral data layer other skills build on.

Contract

Role: Advisory analyst and data layer. Extracts, normalizes, detects anomalies, and scores propensity; does not change the site or run experiments. All statistics and models run in deterministic Python scripts.

Mode:

• quick — normalized metrics + anomaly detection.
• standard (default) — add navigation path and content-affinity analysis.
• deep — add predictive audiences and page-performance correlation.

When to use: GA4/traffic analysis, anomaly detection, session/path behavior, predictive audiences, landing-page performance.

When NOT to use: organic-search keyword/ranking work → seo-content; funnel-step drop-off → funnel-analysis; clustering customers → audience-segmentation. See ../../references/skill-index.md.

Evidence required (inputs):

• workspace/raw/ga4_reports.csv — date, source, medium, page_path, sessions, conversions. Required. If absent, STOP and run data-extraction first (or use the google-analytics-mcp server / scripts/extract_ga4.py to land it).
• workspace/raw/events.csv — event-level data (optional; enables path analysis).

Depends on: data-extraction. Feeds into: funnel-analysis, seo-content, paid-media, audience-segmentation, experimentation, reporting. Builder detail in references/authoring-notes.md.

Hard STOP (analysis validity): anomaly detection requires ≥8 weeks of history for a stable seasonal baseline. If history is shorter, STOP the anomaly step, report descriptive trends only, and label the seasonal baseline as unavailable — do not emit anomaly flags from an unstable decomposition.

Parse $ARGUMENTS for inline paths, date range, or Z-score threshold overrides.

Workflow

1. Validate inputs. Load ga4_reports.csv and/or events.csv; verify required columns. Normalize UTM parameters (lowercase source/medium, trim, URL-decode).

2. Extract-source gate (AskUserQuestion). When fresh data is needed and both a live source and a stale file exist, ask before pulling:

   • Question: "How should I source GA4 data?"
   • Options: (a) Use the workspace CSV as-is — fastest, may be stale; (b) Pull fresh via google-analytics-mcp — when the MCP server is connected; (c) Run scripts/extract_ga4.py — Data API fallback with named dimensions. Skip when the workspace file is fresh and sufficient.

3. Normalize metrics. Aggregate to daily traffic/engagement/conversion; write workspace/processed/web_metrics.json.

4. Anomaly detection. Run scripts/web_anomaly_detection.py: STL seasonal decomposition (period=7) then Z-score residual flags. Apply the Hard STOP gate on history length first.

5. Root-cause decomposition. For each anomaly, break the deviation down by source/medium, device, geography, and landing page; suppress known events (holidays, launches) when a suppression calendar is provided.

6. Navigation paths. Run scripts/path_analysis.py: second-order (bigram) Markov transition matrices; top conversion paths, loops, dead-ends.

7. Content affinity & exit pages. Conversion-lift ratio P(convert|category)/P(convert), ranked; exit-rate ranking weighted by conversion proximity. See references/behavioral_patterns.md.

8. Predictive audiences (deep mode). Run scripts/predictive_scoring.py: logistic regression for convert/churn propensity on behavioral features; temporal holdout validation (target AUC > 0.70).

9. Page performance (deep mode). Correlate load time with bounce rate, segmented by device.

10. Write outputs.

Output Format

Artifacts:

File	Contents
workspace/processed/web_metrics.json	Normalized traffic/engagement/conversion metrics
workspace/analysis/web_anomalies.json	Anomalies with root-cause decomposition
workspace/analysis/navigation_paths.json	Common paths with conversion correlation
workspace/analysis/predictive_audiences.json	Convert/churn propensity scores

Report states: history length used, anomaly threshold, path-analysis order, and predictive-model holdout AUC (or that scoring was skipped).

Financial services mode: track regulatory disclosure page views for compliance verification; authenticated investor-portal analytics require PII handling (GDPR, CCPA, SEC); report cookie-consent rate as a data-quality metric; verify required disclosures rendered before conversion events. Any customer-facing output routes through compliance-review.

Completion status:

• DONE — metrics, anomalies, and requested behavioral artifacts written.
• DONE_WITH_CONCERNS — e.g., short history (anomalies skipped), missing events (no path analysis), or low predictive AUC.
• BLOCKED — no GA4 data and extraction unavailable; state the fix.
• NEEDS_CONTEXT — extraction-source choice unresolved by the user.

Anti-Patterns

• Emitting anomaly flags from fewer than 8 weeks of history.
• First-order Markov chains where bigram transitions are specified.
• Random (non-temporal) holdouts in predictive scoring — they leak the future.
• Skipping UTM normalization, so Google and google split traffic.
• Reporting raw exit rate without conversion-proximity weighting.
• Letting the model estimate anomaly thresholds or propensity scores instead of the scripts.