Analytics & weekly review

Analytics in Zubby are designed for one purpose: make it obvious whether the AI is moving the needle for your store, and where to invest next. This page covers the dashboards, how attribution works, and the weekly review snapshot - the single best read of the week.

The Analytics dashboard

Open Analytics from the sidebar. The page covers the last 30 days by default and leads with a headline KPI grid:

AI-Attributed Revenue - revenue from orders the agent influenced (see attribution below).
Conversation Volume - chats started in the period.
Cart Rescue Rate - conversion rate of triggered cart rescues.
Recovery Conversion - conversion rate of recovery emails sent.

A second Customer health row tracks average order value, cart abandonment rate, return rate, handoff rate, and CSAT. Below the KPIs come the charts and panels: a revenue trend chart, an event breakdown, rescue and recovery funnels, the opening-message A/B experiment (assignments, add-to-cart and lead rates per variant), smart-offer popup performance (impressions, CTR, dismiss rate, trigger mix), back-in-stock and price-drop demand signals, shopper behavior patterns, and the top shopper questions asked in chat.

Conversion funnel

Analytics → Funnel (at /app/<storeId>/analytics/funnel) charts the real AI-assisted path: AI chat opened → product clicked → added to cart (AI) → checkout completed. Each bar shows the absolute event count, the step-over-step conversion rate, and the share of the top of the funnel, plus a headline chat→checkout percentage. A date-range selector offers 7 / 14 / 30 / 90-day presets and a custom range.

Attribution

The headline revenue number is last-touch attribution: an order is credited to the AI (source ai_last_touch) when a product in the order was recommended to that shopper in chat within the 30 minutes before the order was placed. Everything else counts as organic under this reporting rule. A match credits the full order, not just the recommended line. It can over-credit the AI because some shoppers would have bought anyway; missing identity or events can also omit influenced orders. It is not incremental lift.

Interpreting a holdout comparison

For incrementality, Zubby ships an opt-in attribution holdout: a deterministic fraction of shopper sessions can be held out of AI exposure, and conversion in the exposed group minus the holdout group is a directional estimate. The holdout suppresses the chat assistant, but standalone promotions and recommendation surfaces remain visible. The current report requires Shopify cart-token linking, at least 100 visitors per arm and a linked order; it does not establish statistical significance or profit. The holdout percentage defaults to 0, so nothing changes until you turn it on.

Weekly review

Analytics → Weekly Review is the Monday digest: how the agent performed over the last 7 days versus the platform average. It benchmarks three metrics - cart rescue rate, AI conversion rate, and revenue per chat - each with the platform average and your percentile. A snapshot is saved every Monday and emailed to the store owner; the history table keeps the last 12 weeks so you can scroll back through past snapshots. Use Send digest now to trigger the email on demand (and verify the pipeline), or Export PDF for a shareable copy.

Internal comparisons are not an audited benchmark

Platform averages and percentiles depend on the stores, sample sizes, windows and eligibility included in the snapshot. No statistically reviewed cross-customer benchmark is established here. Compare your own like-for-like periods first, and check coverage before drawing conclusions from a percentile.

A weekly reading routine that works:

Open Weekly Review on Monday and check whether the three benchmark metrics moved against last week’s snapshot.
If cart rescue or AI conversion dipped, open the conversations from that week and sample-read what changed.
Fix the root cause where it lives - knowledge gaps in the knowledge base, weak product data in the catalog.

Cohort retention

Analytics → Cohorts (at /app/<storeId>/analytics/cohorts) shows a monthly retention grid: each row is the month a shopper first ordered, each column is months 0–5 since, and each cell is the share of that cohort who ordered again - a quick read on whether repeat purchasing is improving.

Exports

The Export CSV button at the top of the Analytics page downloads the analytics dataset for the store. For prebuilt one-click CSVs (orders, customers, conversations, leads, carts, feedback, AI revenue), use the Reports page. For programmatic access, see the REST API; for a continuous stream of events, see Webhooks.

Comparing with GA4 / Meta CAPI

Zubby also pushes conversion events to GA4 and Meta CAPI when connectors are configured. Attributed credit can differ because the selected report, attribution model, lookback window and identity coverage differ. GA4 is not universally session-based. Record the report and model you selected; see Google’s attribution settings guidance. Reconcile raw transaction IDs, currency, order totals, timestamps, consent, missing events and duplicates before interpreting model differences.

Worked reconciliation example (illustrative)

Choose one paid test order, for example TEST-1042, and the same date range and currency in each system.
Confirm that the order appears once in the commerce platform and once in each configured conversion destination. Check the transaction ID and value before comparing channel credit.
In Zubby, inspect the linked conversation and matching recommended product within the 30-minute window. A chat alone does not qualify.
If the raw order matches but credited channels differ, document the reporting models and windows. If the event is absent or duplicated, investigate ingestion or consent instead of dismissing it as attribution.

Where to dig deeper

Conversations - scoring methodology.
Recovery - attribution for journey messages.
Customers - lifecycle stages and segments behind the health KPIs.