# Why Attribution Reports Disagree: Reconcile the Difference Before Blaming the Model

Follow ten fictional orders from a $910 platform report to $270 of cross-channel search credit. Separate duplicate events, test orders, date boundaries, refunds and credit rules before deciding what needs fixing.

Source: https://dtcup.com/blogs/insights/reconcile-attribution-report-differences/

DTCUP / FIELD NOTES / ACQUIRE

# Why Attribution Reports Disagree: Reconcile the Difference Before Blaming the Model

Follow ten fictional orders from a $910 platform report to $270 of cross-channel search credit. Separate duplicate events, test orders, date boundaries, refunds and credit rules before deciding what needs fixing.

Published Oct 2, 202624 min read

 ![Jack Hu](https://dtcup.com/wp-content/plugins/dtcup-knowledge/assets/jack-hu-small.jpg)**Jack Hu**7 years of experience building and operating brand websites, specializing in website development, paid advertising and SEO. Built a brand website from launch to US$10 million in annual sales.



  On this page  ## The Short Answer

Write the contract for each report, establish an eligible order cohort, then align event quality, date basis and value definition before comparing credit rules. Keep missing observation and unexplained residuals visible. A reconciliation explains reporting differences; it does not establish causal marketing lift.

## Key takeaways

- Report names and matching date pickers do not guarantee matching definitions.
- Duplicate purchases and test orders are errors, not attribution choices.
- Interaction-date and purchase-date cohorts can move orders in both directions.
- Refund treatment must use a defined value basis and cohort clock.
- Different credit rules can reassign revenue without changing business sales.
- Unknown channel history belongs in the outcome total and should not be guessed away.

![Conceptual overhead arrangement of forest-green and sage report strips linked to an ivory ledger card, with a transparent ruler and orange correction tabs.](https://dtcup.com/wp-content/uploads/2026/10/dtcup-a02-operating-guides-20261002-1024x576.webp)Your advertising report says paid search produced 910 dollars. Your cross-channel report says 270. The store reports 880 dollars across all channels. It is tempting to declare one tool wrong, blame attribution and move on. But those three numbers may describe different events, dates, values and credit rules. They may also contain a real implementation error. You need a way to separate the two.

I would reconcile the reports in layers. First establish which business outcomes belong in the comparison. Then align the money definition and date basis. Next inspect observation and eligibility differences. Only after those foundations match should you compare how credit is assigned. Changing attribution models before checking the underlying orders can make a tracking defect look like a philosophical disagreement.

This guide uses a fictional brand, Northbank Goods, and ten invented order records. It contains both expected definition differences and genuine errors so the reconciliation has something concrete to diagnose. The numbers do not come from DTCUP, a client or a named advertising platform. The final bridge will explain the 640-dollar difference without pretending that every real platform provides order-level evidence for all reported conversions.

For choosing a model according to the decision you want to make, read [the attribution decision guide](https://dtcup.com/blogs/insights/choose-attribution-model-channel-decision/). Here the question is narrower: before comparing performance, can we explain why these two reports produce different totals from their particular rules and observations?

## 1. Write down what each report is actually counting

A report name is not a measurement definition. “Revenue,” “purchase value” and “attributed sales” can refer to different amounts. “Conversions” may include several actions or a counting rule that differs from the order system. Before exporting data, write a short contract for each view: the business event, included records, money basis, date basis, time zone, attribution window, credit rule and reporting timestamp.

For Northbank, the business comparison covers purchases made from May 1 through May 7 in the store’s local time. Eligible purchases exclude test orders. Revenue means merchandise value after discounts and after the linked refunds known by the review date, excluding shipping and tax. All values use one currency. The review occurs after the example’s relevant records have arrived.

The fictional platform report instead groups credited purchase events by paid-search interaction date. It uses merchandise value after discounts but before later refunds. Its displayed total includes a duplicated purchase event and a test order because the example implementation failed to exclude them. Those two items are defects, not legitimate attribution choices.

The cross-channel report groups eligible orders by purchase date, applies the reviewed refund amounts and credits the last eligible non-direct channel within a seven-day lookback. Unknown history remains unknown. This is an explicitly defined illustrative rule, not a statement about every tool’s default model or DTCUP’s current account settings.

The store ledger counts the same eligible purchase-date cohort and money basis as the cross-channel report, but it does not assign channel credit. It supplies the outcome total to reconcile against. It is authoritative for the order records within the defined scope; that does not mean it independently knows every advertising interaction that preceded them.

Record the exact column and filters, not just a screenshot of the headline card. A platform’s all-conversions column can differ from its primary-conversions column. A report can be filtered to new customers, a market, a campaign subset or a particular conversion action. If the filters differ, matching the date picker alone will not produce a meaningful comparison.

Google’s documentation on [Analytics-derived conversions in Google Ads](https://support.google.com/google-ads/answer/2375435?hl=en) describes several reasons totals can differ even with a correct setup, including transaction-date treatment, counting settings and import delay. That is a reason to inspect the report contract carefully, not a reason to dismiss every discrepancy as normal.

[![Three fictional report contracts distinguish platform interaction-date gross credited events, cross-channel purchase-date net credited orders and the store's purchase-date net order ledger.](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/A02-contracts.png)](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/A02-contracts.png)The same date range label can conceal different event and value definitions.## 2. Build the outcome ledger before explaining the attribution

Northbank’s ten teaching records are deliberately small enough to inspect. Order O-01 is purchased on May 1 for 100 dollars, after a paid-search interaction on April 29. O-02 is purchased on May 2 for 120 dollars; it has a paid-search interaction on May 1 and a later paid-social interaction before purchase. O-03 is purchased on May 3 for 80 dollars after paid search on May 2.

O-04 is purchased on May 4 for 150 dollars. Paid search occurs on May 3, followed by an eligible email interaction before purchase. O-05 is purchased on May 5 for 200 dollars after paid search on May 4 and a later eligible paid-social interaction. A linked 40-dollar merchandise refund is known at review time, leaving 160 dollars under our net cohort definition.

O-06 is purchased on May 6 for 90 dollars after paid search on May 5. Its purchase event was accidentally sent twice to the fictional platform with ineffective duplicate protection. The business has one 90-dollar order, not two. O-07 is purchased on May 7 for 110 dollars, but no eligible channel history is available in the cross-channel dataset.

O-08 is purchased on May 8 for 130 dollars after paid search on May 7. It belongs to the platform’s May 1–7 interaction cohort but falls outside the store’s May 1–7 purchase cohort. O-09 is purchased on May 4 for 70 dollars. Its old paid-search interaction on April 15 is outside the example’s seven-day eligibility window; a later organic-search interaction receives cross-channel credit.

O-10 is a 50-dollar test order made on May 4 after a May 3 paid-search interaction. It should not belong to the eligible business outcome set, but the fictional platform implementation recorded it. Keep it visible as an excluded record rather than silently deleting the evidence of the problem from the reconciliation.

The eligible purchase-date cohort is O-01 through O-07 plus O-09. Their merchandise value before the reviewed refund totals 920 dollars. Subtracting O-05’s 40-dollar refund gives 880. O-08 is excluded because of the purchase-date boundary; O-10 is excluded because it is a test. The duplicated O-06 event never creates a second business order.

This ledger establishes the business outcome denominator. It does not yet prove which channel deserves credit. It also does not prove that all order events were observed correctly by every platform. That separation is the point: first know which purchases and amounts you intend to explain, then investigate how each reporting system represents them.

[![Fictional purchase cohort totals 920 dollars before a 40-dollar refund and 880 after it; May 8 order O-08 and test order O-10 are excluded, while duplicate O-06 events remain one order.](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/A02-ledger.png)](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/A02-ledger.png)The order ledger defines outcomes. It does not manufacture missing channel history.## 3. Separate implementation errors from legitimate definition differences

The platform’s 910-dollar total includes O-02 at 120, O-03 at 80, O-04 at 150, O-05 at 200, O-06 twice at 90 each, O-08 at 130 and O-10 at 50. O-01’s paid-search interaction is outside the selected interaction-date range. O-07 lacks a qualifying observed search interaction, and O-09’s old search touch is ineligible under the example’s window.

Start by removing the extra 90-dollar O-06 event from the analytical comparison. That leaves 820. Remove the 50-dollar test order, leaving 770. These are data-quality corrections. A different attribution model would not make the duplicate a real order or turn a test transaction into a customer purchase.

In a live implementation, diagnose why the duplicate occurred. The browser and server may have sent the same event without the intended shared identifier, a retry may have used a new identifier, or two integrations may have independently recorded the purchase. Inspect the actual event and transaction handling. Do not assume a particular vendor’s deduplication behavior without checking its documentation and your configuration.

Correct the source of the error as well as the analysis. A reconciliation worksheet that subtracts duplicates every week while the bidding system continues consuming them leaves the operational problem intact. Depending on the platform and supported adjustment process, historical correction may be limited. Record what was repaired prospectively and what remains in historical reporting.

Test-order exclusion also needs a defined rule. A staff email address is not always enough to identify a test, and a zero-value order is not necessarily invalid. Use approved flags or a documented test environment where possible. Preserve legitimate promotional, replacement or other nonstandard business outcomes according to the measurement definition rather than excluding everything unusual.

Check whether a discrepancy is really a value duplication or a count duplication. Two events may share a transaction but carry different item or shipping values. A platform might deduplicate one event type while another remains. Reconcile stable identifiers, event types and values together; matching only the number of rows can conceal the money error.

A useful classification is “confirmed error,” “expected definition difference,” “timing or maturity difference” and “unexplained.” Keep the fourth category. The temptation to force every residual into attribution is strong, especially when stakeholders want a final number. An unresolved discrepancy with an owner and next check is more honest than a neat bridge built from unsupported guesses.

## 4. Align the date basis before comparing the selected week

The cleaned platform subtotal is 770 dollars on its interaction-date basis. O-08 contributes 130 because its paid-search interaction occurred on May 7, even though the purchase happened on May 8. Remove it when moving to the May 1–7 purchase cohort. The subtotal becomes 640.

O-01 moves in the other direction. Its purchase occurred on May 1, so it belongs to the business outcome cohort, but its paid-search interaction was on April 29. Add its 100 dollars when translating the example’s eligible search credit to purchase date. The subtotal becomes 740.

Neither move changes when the customer bought or how much the order was worth. We changed the report’s grouping basis. That is why comparing two identical date-picker labels can be misleading: one selects interactions during the period and follows their conversions, while the other selects purchases during the period and looks back at earlier interactions.

Use the platform’s documented conversion-time columns or equivalent view when available. [Google Ads’ attribution-report documentation](https://support.google.com/google-ads/answer/1722023?hl=en) distinguishes event-time treatment across report types and describes conversion-time columns for aligning the reporting basis. Verify the exact report and column rather than assuming every screen uses the same clock.

Time zones create a separate boundary issue. An order just after midnight in one system can fall on the previous calendar day in another. Store the original timestamp and time zone, then derive the common reporting day consistently. Do not solve the mismatch by manually moving a few inconvenient orders until totals happen to match.

Reporting maturity matters too. A click-date cohort can continue receiving later conversions within its eligible window. An import can arrive after the initial report was exported. A refund can occur after the purchase. Record the “as of” time for every comparison and revisit a stable cohort after the agreed maturation period.

A mature cohort is not one that can never change. Late corrections and supported adjustments may still occur. The practical goal is to choose a review window suitable for the buying cycle and data process, then label later restatements. Comparing a fresh two-day platform cohort with a settled monthly order ledger can make a normal delay look like a tracking outage.

## 5. Align the value definition before assigning channel credit

The purchase-date subtotal for eligible search-associated orders is now 740 dollars before the reviewed refund. Subtract O-05’s 40-dollar merchandise refund to reach 700. We have aligned the date cohort and the value basis, but we have not yet changed the attribution rule.

This example intentionally keeps shipping, tax and currency out of the arithmetic so the reconciliation remains readable. In a real report, check each explicitly. One conversion event might send the full checkout total while the business report uses merchandise revenue after discounts. A difference in value definition can affect ROAS even when both systems count exactly the same orders.

Refund treatment needs a clock of its own. Our report assigns the known refund back to the original purchase cohort. A cash-flow report might instead show refunds on the date they were processed. Both can be useful, but they answer different questions. Do not subtract a month’s refund cash movements from an unrelated purchase cohort and call the result net attributed revenue.

Partial refunds must be linked at the appropriate level. O-05 retains 160 dollars, so any full-order channel credit under the example’s rules should use 160 on the net basis. If a report allocates item-level credit or contains multiple currencies, the reconciliation may need more detail. Avoid assuming that an order-level subtraction distributes correctly across every item or campaign view.

Keep the original event value and the analytical normalized value as separate fields. That allows a reviewer to understand what was actually sent and what the business comparison uses. If the platform supports a value adjustment, record whether it was submitted and accepted. Do not claim that a local spreadsheet correction changed the platform’s stored conversion value.

At this point, 700 dollars is still not the cross-channel paid-search result. It represents the eligible purchase cohort’s net value among orders associated with the example’s qualifying paid-search interactions. The remaining difference concerns credit assignment, which we can now examine without mixing it with duplicate events, tests, date boundaries or refunds.

## 6. Explain the remaining difference through the credit rule

The fictional platform credits qualifying paid-search interactions in its own reporting scope. The cross-channel report uses the last eligible non-direct interaction across its observed channels. That moves three orders away from search: O-02’s 120 dollars go to paid social, O-04’s 150 go to email, and O-05’s net 160 go to paid social. Together they account for 430 dollars.

Subtract 430 from the aligned 700-dollar search-associated subtotal and the cross-channel search result is 270. That consists of O-01 at 100, O-03 at 80 and O-06 at 90. The full bridge from 910 to 270 is now explained: minus 90 duplicate, minus 50 test, minus 130 outside the purchase period, plus 100 entering on purchase date, minus 40 refund and minus 430 reassigned by the cross-channel rule.

The 430 dollars did not disappear from the store. It moved to other channels under a different allocation rule. The complete cross-channel total is paid search 270, paid social 280, email 150, organic search 70 and unknown 110. These add to the store’s 880-dollar net outcome total for the defined cohort.

This reconciliation does not prove that the last channel caused the purchase. Paid search may have helped O-02, O-04 and O-05 even when it receives no final credit in the illustrative model. The platform’s claim also does not prove that the orders would not have happened without search. We have explained reporting differences, not estimated incremental sales.

Do not add the platform’s search credit to the cross-channel social and email amounts. Those views overlap on the same business outcomes. Summing them would count parts of the same orders more than once. Choose one coherent allocation for a cross-channel total and keep platform-specific views as separate diagnostic lenses.

Fractional models create another legitimate difference. If a defined model distributes an order across multiple eligible interactions, a channel can receive a fraction of an order count or value. That does not necessarily mean the tracking system created part of a transaction. Inspect whether fractions sum correctly within the model’s eligible scope, allowing for displayed rounding, before treating them as a data defect.

Be precise about what the bridge establishes. In our fictional data, every adjustment is supported by the constructed ledger and rules. In a real platform, some modeled or aggregated credit may not be traceable to an exported order identifier. You can still reconcile definitions and observable subsets, but you should not invent order-level explanations for the remainder.

[![Reconciliation moves from 910 dollars through duplicate, test, purchase-date and refund adjustments to 700, then subtracts 430 of cross-channel credit reassignment to reach 270.](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/A02-bridge.png)](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/A02-bridge.png)The bridge explains different layers. Only the first two adjustments are implementation errors in this example.## 7. Distinguish missing observation from a different allocation

O-07 has 110 dollars of eligible net revenue but no usable channel history in the cross-channel dataset. It stays in the outcome total and appears as unknown. This is different from O-04, whose observed path exists but receives email credit under the chosen rule. One is an observation gap; the other is an allocation choice.

Unknown history can arise from many causes: missing campaign parameters, unavailable identity continuity, consent choices, redirects, an integration gap or a purchase path the system does not observe. The report should not assign a cause without evidence. A blank source field alone cannot tell you which mechanism failed.

Keep the unknown amount visible when evaluating coverage. If the report silently drops O-07, its channel totals would sum to 770 instead of the store’s 880. The apparent reconciliation might look cleaner because only attributable orders remain, but the business denominator has changed. Report both the attributed share and the total eligible outcome set.

A platform may report conversions it can estimate or associate through mechanisms unavailable in your first-party path dataset. That does not automatically prove the platform is wrong or prove the estimate is exact. Record the reporting type and compare what can be observed directly. Do not force aggregate modeled values into fictional individual customer journeys.

Likewise, server-side event collection does not automatically create a complete cross-device identity history. It can improve the reliability of particular events while leaving gaps in prior interactions or permissions. Keep event delivery, order matching and journey identity as separate checks. A successfully received purchase event may have no eligible preceding channel in the receiving system.

Review a small set of unexplained orders through permitted evidence. Confirm the purchase exists, the expected event was generated, the identifier was stable, the destination accepted the event and the relevant reporting delay has passed. This sequence can isolate a delivery problem without speculating about attribution. Avoid collecting more personal data than the investigation requires.

When only aggregate data is available, choose aggregate checks. Compare daily or cohort totals after aligning definitions, examine changes around a deployment and check whether the discrepancy is concentrated in a device, market or payment path. These patterns can guide investigation, but they do not by themselves establish the cause of every missing order.

## 8. Check interaction eligibility and report scope explicitly

O-09’s April 15 search interaction is outside the fictional seven-day window for its May 4 purchase. That exclusion is a rule, not a statement that the earlier ad had no influence. The later organic interaction receives credit in the example’s cross-channel model because it is eligible and last among the observed non-direct interactions.

Changing a window can change which interactions qualify and sometimes which outcomes appear in a platform view. Keep the conversion cohort fixed when you want to compare allocation sensitivity. If you change the cohort and the window simultaneously, the difference mixes more than one effect and becomes harder to interpret.

Distinguish click-based and view-based eligibility where the report supports them. A site analytics dataset may record a tagged visit but not an advertising impression. Comparing it with a platform view that includes qualifying impressions can create an expected observation and rule difference. Verify the actual setting; do not assume a universal default or treat every ad exposure as a directly observed individual visit.

Campaign and network coverage can differ across reports within the same platform. A specialized attribution view may not include every conversion source that appears in a broader performance report. Check the current documentation and account configuration for the exact view. The fact that both screens carry the same platform logo does not make their eligible data identical.

Conversion action selection also changes scope. A report that includes purchases, newsletter signups and leads should not be compared with an order count. Even two purchase actions may represent the same business event collected through different implementations. Inspect their purpose, primary or secondary status where applicable, and whether the reporting column combines them.

Do not change production conversion settings merely to make two exported reports match. Those settings may affect optimization as well as reporting. First reproduce a comparison in a controlled analytical view, document the reason for any proposed correction and understand the operational consequence. Reconciliation is an investigation, not permission to rewrite the measurement system until the numbers agree.

Maintain a distinction between historical rules and current rules. A model or window change may affect a report’s presentation differently from its effect on previously recorded data. Verify the platform’s actual behavior rather than assuming the setting retroactively rewrites every historical result. Record the effective date and keep comparisons consistent around the change.

## 9. Build a reconciliation table that another person can challenge

The useful artifact is not a screenshot with an annotation saying “attribution difference.” It is a table with a starting amount, each adjustment, the rule being aligned, supporting evidence, the resulting subtotal and a status. A reviewer should be able to challenge one adjustment without rebuilding the entire analysis.

For the Northbank example, label the duplicate and test exclusions as confirmed implementation errors. Label the O-08 removal and O-01 addition as date-basis normalization. Label the refund as value-basis normalization. Label the final 430-dollar change as credit-rule reassignment. These categories point to different owners and different next actions.

Make the adjustment sequence explicit. A refund on an order already excluded from the purchase cohort should not be subtracted again later. A duplicate test event can satisfy two error descriptions, but it must not be removed twice. The teaching example keeps the adjustments distinct; a real ledger needs record-level flags or a clear precedence rule to prevent overlapping explanations.

Keep intermediate sets when possible. After each step, record which orders or events remain and how their values were normalized. This allows a reviewer to see whether a subtotal changed because records were removed, values changed or credit moved. A list of plausible reasons without a reproducible mapping is not a reconciliation.

Use stable identifiers for matching, and inspect exceptions rather than forcing them. A missing transaction identifier, merged order, split payment or platform aggregation can prevent a simple one-to-one join. Decide how those records should be represented under the business definition. Do not treat a fuzzy title, amount and timestamp match as conclusive identity without an appropriate review.

Record unresolved residuals by type and magnitude. Some may be expected because the source only exposes modeled aggregate data. Others may be a suspected event-delivery gap or a filter you cannot yet verify. Assign a next check and an owner. The goal is to reduce uncertainty honestly, not to force the residual to zero at any cost.

Save the reporting timestamp and configuration evidence with the table. A colleague re-running the report a week later may see a different total because conversions matured or refunds arrived. Without an as-of record, that normal change can look like a calculation mistake. Preserve both versions and explain the restatement.

Retain enough numerical precision to reproduce totals. Fractional credit can create small displayed differences when each row is rounded before summing. Calculate with the underlying precision and round for presentation, then identify any remaining rounding residual explicitly. A one-cent display discrepancy should not trigger the same investigation as a missing purchase, but it should not be disguised as a channel adjustment either.

Give the table a simple independent check: the final channel amounts, including unknown, must equal the eligible outcome total under a full-allocation model. If the selected report intentionally covers only a subset, state that denominator instead. This check catches omitted records and accidental duplication without making any claim that the chosen attribution model measures causality.

[![Reconciliation checklist requires an adjustment amount, definition layer, supporting records, resulting subtotal, execution status and owner for unresolved differences.](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/A02-checks.png)](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/A02-checks.png)A reason is not a quantified explanation until it is tied to the affected records or a documented aggregate limitation.## 10. Investigate responsibly when order-level matching is unavailable

The Northbank example gives us unusually complete evidence so the logic is visible. Real investigations often start with one order ledger, one aggregate platform export and an incomplete event log. You may not be able to reproduce every platform conversion as a row in your own system. The correct response is to define a smaller verified comparison, not to claim the same level of certainty as the worked example.

Begin with the simplest invariant you can test. Does the website record one eligible purchase event for each authorized test transaction? Does the value match the documented merchandise basis? Does the event carry the expected currency and transaction identifier? These checks can establish whether the collection path behaves as intended without needing access to a platform’s entire attribution mechanism.

Then inspect the delivery boundary. A request sent from the browser is not the same as an event accepted by the destination, and acceptance is not the same as inclusion in a particular report. Preserve the supported response and status evidence. If the destination rejects a field or the integration queues a retry, that helps explain a collection issue before you look at channel credit.

Use an approved test process that avoids polluting production optimization. The business should know which test records were created and how they are identified or excluded. Do not create a large set of fake purchases in a live conversion action merely to investigate a mismatch. A small, controlled verification with known identifiers is more useful than additional unexplained events.

For live data, compare stable daily totals after the relevant processing delay. Look for whether the gap is roughly proportional to volume, concentrated on specific dates or introduced suddenly after a deployment. A constant percentage gap may suggest one kind of coverage issue; a sharp break may suggest a configuration or code change. These are investigation clues, not proof of cause.

Break down the comparison only where the dimensions are genuinely comparable. Device labels, market definitions and channel names can differ across systems. A mobile-browser subset in one export may not match a platform’s cross-device conversion classification. Write the mapping before comparing the segments, and retain an unmapped category when a reliable relationship is unavailable.

Pay special attention to checkout paths that differ from the main journey. A returning customer, accelerated payment method, app purchase or separate checkout domain may trigger a different event sequence. The goal is to identify a reproducible path with a missing or duplicated event. Do not assume that a successful desktop test proves every purchasing route works.

Inspect recent changes with a timeline. Note deployments, consent-interface changes, tag updates, payment changes, domain redirects and conversion-setting changes. Compare the discrepancy before and after each relevant event while accounting for ordinary sales variation and reporting lag. A temporal association narrows the search, but the repair still needs a mechanism that explains the observed behavior.

If a vendor support case is necessary, provide the smallest permitted example that demonstrates the mismatch. Include the report name, column, filters, time zone, conversion action, reporting timestamp and expected behavior. Use redacted or authorized test identifiers as appropriate. A message saying “your numbers are wrong” gives the receiving team little to investigate and can lead to generic explanations that do not address the actual case.

Set a stopping rule for the investigation. Some differences cannot be decomposed fully because the platform exposes only aggregate or modeled outputs. Once documented definition differences and observable collection checks are complete, label the remaining amount according to what is known. Continuing to invent increasingly elaborate explanations can be less useful than deciding how the business will use the report despite its limitation.

That decision can be explicit: use the order ledger for business totals, the normalized cross-channel view for a consistent observed allocation, and the platform report for its own campaign diagnostics. Retain the residual as a measurement limitation and monitor whether it changes materially. This is a practical operating choice, not a declaration that the platform estimate is either fraudulent or unquestionably correct.

Do not distribute an unexplained residual proportionally across known channels just to make a chart sum to the store total unless you clearly label that as an allocation assumption. Doing so creates apparent precision without new evidence. If stakeholders need a scenario, show both the observed allocation and the assumed redistribution so they can see which conclusion depends on the assumption.

When comparing periods, monitor the stability of coverage as well as the reported channel result. Suppose unknown revenue grows sharply while paid-search credit falls. That could affect interpretation even if the attribution rule did not change. It does not prove search lost those orders, but it means the apparent performance movement may be entangled with a measurement change.

A final useful check is whether the recommended business action survives the uncertainty. If a budget decision flips only after assigning every unexplained order to one favored channel, the evidence is weak. If the decision is supported across reasonable interpretations and by contribution economics or a suitable experiment, the unresolved reporting gap may be less decisive. Explain that sensitivity instead of hiding it.

Keep the distinction between an analytical workaround and a system repair. A normalized export can make reporting usable while an event issue is investigated, but it does not change the signal used by automated bidding. A repaired event pipeline may improve future data without making old cohorts complete. State which layer changed and verify that layer rather than reporting a universal fix.

## 11. Turn each finding into the right operational action

Different findings should produce different repairs. A duplicate purchase event needs an implementation fix and a supported historical correction if available. A test-order leak needs a better test and exclusion process. A date-basis difference needs a consistent comparison view. A credit-rule difference needs a clear decision about which lens serves the business question.

A refund mismatch may require both a reporting definition and an integration change. If the business wants net cohort revenue but the platform receives only initial purchase value, decide whether supported adjustments should be sent and how their effect will be monitored. Do not assume a platform report has netted refunds simply because the store ledger has.

An unknown-history gap needs investigation of coverage and collection, not automatic reassignment to paid media. A reporting delay may require waiting for the agreed maturity window. A genuinely unexplained residual may require a vendor support case with a safe, reproducible example. Avoid making one tool the universal owner of every mismatch.

Verify the repair with a small authorized test and then a live observation window appropriate to the change. For duplicate handling, confirm one business purchase creates the intended number of accepted events and the correct value. For date alignment, check boundary cases on both sides of midnight or the reporting period. For a refund adjustment, confirm the destination’s accepted result and report behavior.

Keep the original mismatch in the change record. State the before and after behavior, what was actually changed, what was only normalized in analysis and what remains unresolved. A clean final chart should not erase the evidence that the bidding input or reporting process was wrong for a period.

Do not use the reconciled 270-dollar search figure alone to decide whether to cut search spending. It is the result of a defined last-interaction rule over an observed cohort. The platform-associated 700-dollar aligned value supplies a different view of participation, and neither establishes causal lift. Budget decisions need the broader evidence discussed in the attribution decision guide, including economics and suitable experiments.

A useful monthly meeting can therefore hold three views side by side: the business outcome ledger, a coherent cross-channel allocation and platform-specific diagnostic reports. Explain their contracts and avoid summing overlapping claims. The disagreement becomes manageable when each number has a purpose and a known boundary.

[![The fictional 880-dollar net purchase cohort is allocated as paid search 270, paid social 280, email 150, organic search 70 and unknown 110, preserving the complete business total.](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/A02-allocation.png)](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/A02-allocation.png)A coherent allocation preserves the outcome total and keeps unknown history visible.The aim is not to make every platform display the same number. It is to know which differences are errors, which follow from definitions and which remain unexplained. Once that is clear, the team can fix the measurement problems, choose a consistent decision view and stop arguing about totals that were never designed to mean the same thing.



 ## Sources

1. [support.google.com — 2375435](https://support.google.com/google-ads/answer/2375435?hl=en)
2. [support.google.com — 1722023](https://support.google.com/google-ads/answer/1722023?hl=en)

CONTINUE / THE TOPIC

## Keep the next question connected.

[NEXT BEST READ**Run a Marketing Measurement Meeting That Ends with One Useful Decision**](https://dtcup.com/blogs/insights/marketing-measurement-review-decision-method/)[RELATED GUIDE**Which Attribution Model Should You Use? Start with the Decision**](https://dtcup.com/blogs/insights/choose-attribution-model-channel-decision/)[RELATED GUIDE**Choose an Attribution Window Around the Buying Cycle**](https://dtcup.com/blogs/insights/choose-attribution-window-buying-cycle/)[RELATED GUIDE**Email and SMS Attribution: Stop Giving Every Returning Sale to the Last Message**](https://dtcup.com/blogs/insights/email-sms-attribution-holdout-contribution/)

Still have a question?[Ask the DTCUP Community ↗](https://dtcup.com/community/)
