A marketer opens an advertising platform, GA4, and the commerce system. Each shows a different number. The fastest response is often “tracking is broken,” followed by a request to make the reports match. That can hide real defects, but it can also erase legitimate differences in attribution, identity, timing, and outcome status. Diagnosis begins by accepting that equality is not the default.
Build a discrepancy worksheet
Record the report name, source, property or account, outcome, event definition, date field, reporting date, time zone, currency, attribution window, model, eligible touchpoints, view-through rules, identity, consent, status, filters, and update latency. Capture screenshots or exports with timestamps. Without this worksheet, teams compare labels rather than definitions.
Diagnose in the right order
Outcome and status
Confirm whether reports count platform-attributed conversions, analytics events, valid leads, created orders, paid orders, or net orders after cancellation or refund. Check transaction IDs and test data. A platform can credit an outcome that the commerce system later cancels; analytics can receive a purchase before a refund arrives. These are lifecycle differences, not necessarily lost tracking.
Time and attribution
Align time zones and understand whether the report assigns the outcome to interaction date or conversion date. Compare click and view windows, eligible channels, last-click or data-driven rules, and direct-traffic treatment. A conversion near a date boundary can move between days. Late processing can change recent reports after the first review.
Identity and observability
Document cookies, login, platform identity, cross-device behavior, consent, blockers, app browsers, and offline activity. Advertising platforms may connect interactions that analytics cannot, while analytics may observe channels the platform excludes. Neither sees every influence. Do not call the difference a precise amount of “missing” data.
Implementation and transport
Only after definitions are aligned should the team inspect triggers, tags, data layers, network requests, server events, event IDs, duplicates, parameters, redirects, and platform diagnostics. Compare the expected event with actual source records. A global ratio does not reveal whether one market, browser, event path, or integration is failing.
Use patterns to narrow the cause
A sudden change at a release boundary suggests implementation, consent, domain, or campaign configuration. A stable structural difference suggests models, identity, or status definitions. One-device or one-browser differences suggest client behavior. One payment method may reveal confirmation routing. One campaign may have broken UTMs or an incorrect destination.
Graph the difference and ratio over time, but do not use the ratio as an automatic correction factor. Add annotations for releases, consent changes, attribution-setting changes, domain migrations, checkout updates, and campaign launches. Compare adjacent events to locate where paths diverge.
Inspect duplicates and missing signals separately
For duplicate outcomes, compare transaction or lead identifiers, event IDs, source, timestamps, refreshes, retries, platform/app overlap, and browser/server deduplication. For missing outcomes, sample records and trace expected trigger, data layer, tag, request, consent, response, processing, and report. Do not infer one defect from an aggregate gap alone.
When sampling, choose representative outcomes across devices, markets, sources, and paths. Protect personal data. Record why each sample was selected and avoid presenting a handful of records as a complete prevalence estimate.
Create a reconciliation view
Use the business system as the governed record for its defined outcome and lifecycle. Show analytics-observed outcomes and platform-attributed outcomes as separate perspectives. Include cancellation, refund, qualification, or pipeline status where relevant. State the expected reason reports differ and the data-quality status of each.
For campaign operation, platform attribution may remain useful. For cross-channel and business review, use shared outcomes plus model context. Do not sum attributed conversions from multiple platforms. If leadership needs one number, agree which governed outcome answers that business question rather than averaging several reports.
Resolve the actual problem
If a tag or integration defect is confirmed, define the correction, affected period, acceptance test, owner, and rollback. Mark historical data rather than fabricating a repair when raw events are unavailable. If the difference is definitional, update report labels, dictionaries, and training. If it is an observability limit, state it and adjust decision guardrails.
After change, repeat the worksheet and sample traces. Monitor for a suitable period. Closing a discrepancy means explaining and controlling the difference, not forcing every report to display the same count.
Create a discrepancy diagnostic record
Open an investigation with exact sources, report names, date range, time zones, currencies, attribution settings, filters, conversion definitions, and extraction times. Capture the size and direction of the difference. “The numbers do not match” is not a diagnosis; a statement tied to a device, date, event, and release is testable.
Sample individual conversions where identifiers allow, while protecting personal data. Trace the business action through the governed outcome system, browser or server request, analytics event, and platform receipt. Separate configuration changes from customer behavior before deciding that a technical defect exists.
Use a controlled comparison window
Choose a settled period and avoid mixing partial current-day data with completed days. Preserve raw exports before normalizing time zone or currency in a separate comparison. Compare trends before totals: a stable proportional difference often points to scope or attribution, while a sharp break can point to a release, connector, consent, or definition change.
Communicate the result without forcing a match
Close with one of four outcomes: technical defect, configuration mismatch, expected methodological difference, or insufficient evidence. State which source governs finance, which supports optimization, and which supports journey analysis. Record the action, owner, validation step, and review date.
Keep a discrepancy register for recurring differences, their expected direction, typical cause, affected reports, owner, and review trigger. Escalate when a new pattern falls outside documented behavior or threatens a material decision. Urgent duplicates, missing purchases, exposed data, or broken destinations should be contained while the deeper analysis continues.
Choose samples deliberately. Include different devices, markets, payment or lead paths, consent states, and acquisition sources when those dimensions could change observability. Record why each case was selected and do not present a small diagnostic sample as a prevalence estimate. The purpose of tracing records is to discover where paths diverge, not to manufacture statistical certainty from a handful of outcomes.
Preserve the original exports, screenshots, network evidence, and configuration records with timestamps. Perform transformations in a separate comparison file and document every normalization. If one report uses interaction date and another uses conversion date, show both interpretations before choosing a shared view. This prevents an analyst from “fixing” a discrepancy by silently changing the evidence under review.
Write the conclusion for the decision maker who will use it. Explain whether the difference changes budget optimization, finance reporting, journey analysis, or none of those decisions. Include the limitation in the dashboard or review note where it matters. If evidence remains insufficient, state what additional observation or validation would be needed instead of assigning a confident cause.
Revalidate after the correction using the same event definition, representative paths, and comparison method. Confirm that deduplication, consent behavior, transaction identifiers, and downstream status still behave as intended. A repaired dashboard is not enough if the source event remains wrong, and a repaired source does not justify rewriting historical records that cannot be reconstructed.
Set a review trigger for releases that affect checkout, forms, domains, consent, tag management, server integrations, CRM mapping, or conversion settings. Attach configuration dates to the register so a later analyst can distinguish an expected methodological shift from a fresh defect. This maintenance step turns one investigation into a reusable control without pretending every future discrepancy will have the same cause.
Common discrepancy-diagnosis mistakes
- Comparing similarly named but different outcomes.
- Ignoring interaction-date versus conversion-date reporting.
- Calling all platform credit directly observed conversions.
- Applying a blanket correction ratio.
- Debugging tags before aligning attribution settings.
- Sampling only one browser, market, or payment path.
- Changing dashboards to hide an unresolved source defect.
Reporting discrepancy checklist
- Complete the source, outcome, timing, and model worksheet.
- Align time zones, currencies, filters, and status.
- Document identity, consent, and channel observability.
- Check releases and structural patterns over time.
- Trace duplicates and missing events separately.
- Sample records across representative paths.
- Build a labeled reconciliation view.
- Correct, document, and revalidate the confirmed cause.
