Define the reconciliation contract
Write down the property, stream, reporting time zone, date range, export mode, report surface and metric definitions before running a query. A raw event count is not the same thing as a report's users or sessions. Compare like with like: events to events first, then reconstruct more complex metrics using documented rules.
Separate three categories of variance: implementation defects such as missing events; processing or freshness differences such as late arrivals; and structural differences such as modeled data, reporting identity, attribution or high-cardinality aggregation. Each category requires a different response and should not be hidden under a single acceptable variance percentage.
Check export coverage and freshness
Inventory daily event tables and any intraday or streaming tables. Confirm that expected dates exist, inspect row counts by day and event name, and review when late events are incorporated. Google documents that daily tables can be updated for up to 72 hours after the event date, so avoid treating yesterday's export as immutable.
Verify whether export limits, consent configuration, stream changes or deleted data could affect coverage. Use the property's reporting time zone when comparing event_date with Analytics reports; event_timestamp is stored as a UTC timestamp in microseconds. A date-boundary mismatch can look like missing events when the rows are merely assigned differently.
Inspect event schema and key parameters
For each business-critical event, validate event_name, event_timestamp, user_pseudo_id where available, event_params and relevant ecommerce fields. Use a schema dictionary to define required parameters, accepted types and allowable nulls. Unexpectedly missing transaction_id or form identifiers should be investigated before a dashboard is built on top.
Test for duplicate purchase or lead events using stable business IDs when present. Do not assume distinct event rows equal distinct customers or transactions. A retry, duplicate tag or server/client double-send can inflate totals while appearing valid at the schema level. Build checks for missing identifiers, impossible values and sudden event mix changes.
Reconstruct users and sessions carefully
Users in the interface may refer to active users, not every distinct identifier seen in raw export. Google Signals and reporting identity can also affect UI numbers without exposing equivalent identity data in BigQuery. Never report COUNT(DISTINCT user_pseudo_id) as guaranteed parity with the UI's Users metric.
For sessions, use a documented approach based on the available user identity and ga_session_id rather than counting session_start rows indiscriminately. Check cross-midnight sessions and missing identifiers. Scope matters: event-scoped traffic_source information is not automatically equivalent to session-level acquisition dimensions in standard GA4 reports.
Handle attribution and modeled data honestly
Analytics reporting may include modeled behavior or attribution processing not present in raw event export. The interface can also group high-cardinality dimensions into an other row or apply thresholding. These are product differences, not necessarily data-pipeline failures. Document the exact surface and transformation responsible for any known difference.
If stakeholders need comparable campaign reporting, choose one primary definition for each KPI and label its source. For example, use GA4 standard reports for a marketing dashboard and BigQuery for granular QA, cohort analysis or joins to CRM outcomes. Avoid quietly swapping one definition for another mid-year.
Automate a small, useful parity scorecard
Schedule checks on settled data for event counts by name, missing daily tables, parameter completeness, duplicate business IDs and major shifts in device or stream mix. Establish alert thresholds using normal variability, then attach sample records and a clear owner to each exception. Do not alert on every small variance.
Pair this audit with GA4 channel-grouping governance and the dashboard metric dictionary. Use the UTM Validator to reduce avoidable acquisition-label inconsistencies. A good outcome is trustworthy, explainable measurement—not artificially identical numbers across fundamentally different reporting products.