The decision this guide supports: what measurement and technology foundations are sufficient to make better commercial decisions, and which gaps are materially misleading? Buying another analytics platform will not repair incompatible metric definitions or missing source permissions.

List core systems for orders, products, advertising, site behavior, customer relationships and costs. For each metric, name a source of truth appropriate to its purpose. Payment and order records support financial reconciliation; analytics platforms support behavioral interpretation but often have consent and attribution limitations.

Create a metric and source contract

Document definitions, units, time zones, currency conversion, gross versus net revenue, refund handling, customer identity and attribution windows. Track source account permissions, last successful ingestion, expected latency, known missing intervals and retry status. A dashboard should display stale evidence as stale, not replace it with zero or silently use an old snapshot.

Reconciliation should compare like with like. Shopify order revenue, GA4 purchase events and ad-platform reported conversions answer different questions. Differences can be legitimate, but unexplained drift undermines confidence. Audit what counts as a purchase before interpreting the gap.

Use layered, bounded quality assurance

Separate ingestion checks, identity and duplicate handling, source-to-warehouse comparisons and reporting-model validation. Establish a shared QA framework with connector-specific adapters rather than unrelated testing systems for each vendor. Daily and on-demand runs should produce dated, verifiable outcomes, not a permanent green badge.

Repairs should be triggered by evidence of actual change, with idempotency and bounded concurrency. Avoid marking unchanged duplicate records as dirty if doing so launches expensive downstream transformations. Cost and correctness are connected: a stable design should avoid repeated unnecessary work while detecting real gaps and provider restatements.

Worked example: two platforms disagree on sales

Suppose a hypothetical retailer sees 1,200 paid orders in its commerce system but fewer purchase events in web analytics. An analyst checks event definitions, consent coverage, order cancellations, time zones and deduplication. After explaining the remaining gap, the team may use commerce orders for financial contribution and the behavioral subset for funnel diagnosis, with limitations declared.

Blindly forcing the two totals to match would obscure useful distinctions. The system should present both their provenance and the reconciliation result.

Minimum trustworthy-data checklist

• Define critical metrics with owner, calculation, source and data freshness

• Monitor permission failures, missing dates, duplicate identities and schema changes

• Reconcile source, warehouse and marts on representative windows

• Expose last-good, last-attempt, pass/fail/blocked and estimated completeness

• Bound repairs and record processing cost, duration and outcomes

Self-check: can a business operator tell whether a metric is trustworthy enough for the pending decision without asking a developer to inspect logs? If not, improve the evidence interface before expanding the dashboard. G12 describes how that evidence should be used in experimentation and safe changes.

← Previous guide · Next guide →