Strategy
Part of The honest guide to marketing attribution
Why Marketing Attribution Numbers Change After Close
Why do marketing attribution numbers change after close? Compare late outcomes, model runs, imports, date bases, and modeled values before reading the trend.
When marketing attribution numbers change after a reporting period closes, the difference does not automatically mean customer behavior changed. Reconcile the old and new reports by identifying whether late outcomes, adjustments, model processing, imports, source corrections, or a changed reporting basis explain the restatement. That distinction helps prevent a revised number from being read as a new performance trend.
Why a closed-period report can move
A closed report is a snapshot, so later inputs or processing can still affect its numbers. First identify what changed: the total number of outcomes, the allocation of credit among channels, or the period in which outcomes appear. These are different kinds of restatement and call for different explanations.
| Possible source | What may change | Record for the comparison |
|---|---|---|
| Late outcomes | Conversions are added after close | Arrival date and affected period |
| Invalid-traffic adjustments | Counts are removed after review | Adjustment date and amount removed |
| Model processing | Credit is redistributed between channels | Model version and run date |
| Imports | Delayed or offline records are added | Import date and source file |
| Source corrections | An upstream field is repaired | Reason and corrected field |
A change in channel shares with no corresponding change in the total may point toward model processing, a lookback window, identity coverage, or the report's time basis. A changed total may instead reflect late outcomes, an adjustment, an import, or a source correction. These are clues, not proof, so compare the old and new report details before assigning a cause.
Attribution Failures and Fixes
Failure
- Credit called cause
- Attributed return called impact
- Undefined outcome
- Leads, sales, value blended
- Hidden window
- Paths shorten after change
- Weak identity
- Known users dominate shares
- Immature value
- Recent cohorts look profitable
- Model drift
- History or bids change
- Time mismatch
- Reports group different dates
- Silent correction
- Past decisions retain wrong context
Warning
- Credit called cause
- Use bounded language, test
- Undefined outcome
- Write separate contracts
- Hidden window
- Version and disclose lookback
- Weak identity
- Show coverage, unmatched outcomes
- Immature value
- Use fixed-age closes
- Model drift
- Log version, parallel run
- Time mismatch
- Align or label time basis
- Silent correction
- Notify owners, reassess action
Correction
- Credit called cause
- Undefined outcome
- Hidden window
- Weak identity
- Immature value
- Model drift
- Time mismatch
- Silent correction
Use the figure as a catalog of possible attribution failures, not as a diagnosis of a particular restatement. The relevant question is which setting, input, or processing event changed between the two reports, and whether that change affects totals, channel allocation, or the period in which outcomes appear.
Compare the same report before interpreting the difference
Walk through the old and new versions side by side. Compare dates, outcome definitions, time basis, lookback window, model version, inputs, and identity coverage before assigning a cause.
A short comparison can expose a changed basis that would otherwise look like a channel gain or loss, rather than a change in customer behavior.
- Confirm the reporting period and outcome definition.Check that both reports cover the same dates and count the same outcome. If either differs, label the comparison rather than treating it as a like-for-like change.
- Check the time basis.Reports may group outcomes by different dates. Note which date basis each report uses; otherwise, the same activity can appear in different periods.
- Compare the lookback window.A changed window can shorten or lengthen the paths eligible for credit. Record the setting used in each report.
- Compare model version and run date.If history or bids changed, note the model version and when each report was processed. A changed allocation is not, by itself, evidence that customers changed their behavior.
- Look for new or corrected inputs.Check for late outcomes, invalid-traffic adjustments, imports, and source corrections. Record which period and report they affect.
- Check identity coverage.If known users make up a different share of the reported paths, channel shares may shift. Show coverage and unmatched outcomes rather than treating the observed paths as the entire result.
If a shift remains unexplained after these checks, do not describe it as a market or channel trend. Keep it labeled as unresolved until its source is clear, and preserve the old and new values rather than smoothing over a change whose basis has not been established. That keeps an uncertain difference from being mistaken for evidence of changed customer behavior.
Read conversion lag separately from performance
Google Ads conversion lag reporting estimates how selected metrics may change as delayed conversions arrive. A projected adjustment is not a mature outcome, and it does not prove that a channel caused the conversion.
Treat it as an indication that a recent period may still move, not as a final result.
For a fair comparison, compare cohorts at the same age and mark the newest periods as incomplete. That prevents a recent cohort with less time for delayed outcomes to arrive from being read as worse or better than an older cohort. For the broader question of outcome maturity in campaign reporting, see campaign measurement and the campaign measurement checklist.
Keep modeled values visible in the report
Google Analytics documents modeled key events and their product context. When modeled values appear, label their coverage so readers can distinguish them from directly observed outcomes.
A model label does not validate the outcome definition or establish a causal interpretation, and it should not obscure whether a revised total reflects modeled values, observed inputs, or both.
Modeled Value Correction Steps
- Capture report settings, model, window, filters, extraction time
- Reconcile source outcomes and eligible paths
- Compare observed and modeled fields where available
- Identify reports, bids, budgets, claims affected
- Publish dated correction and material difference
- Add test or approval to prevent recurrence
Use the checklist to make modeled coverage visible when a report is corrected. If the old and new totals differ, state whether the change came from modeled values, observed inputs, or both when that information is available. Do not silently present a revised modeled total as though it were unchanged observed data.
Preserve the history of a restatement
Keep the value that was previously reported, the revised value, when the change was made, and the reason for it.
Also note which decisions were made using the earlier number and whether their context needs review. A transparent series break is more useful than a smooth history assembled from incompatible definitions; record any revisions and corrections alongside the changed values.
For broader checks on interpreting attribution results, see reading attribution numbers without being misled. The core task here is narrower: explain what changed between two versions of a closed-period report before using either one to tell a performance story. Keep the comparison tied to the same dates, outcome definition, time basis, and inputs, and state any unresolved difference plainly.
Common questions
Why can conversions change after the period closes?
Late outcomes, invalid-traffic adjustments, model processing, imports, and source corrections can change a report after close. Comparing the old and revised versions can show which input or processing event changed, and whether the total, channel allocation, or period changed, rather than assuming customer behavior changed.
Does a revised channel split prove performance changed?
No. A changed setting, model run, time basis, or identity coverage can alter channel shares. Compare the report details before treating the shift as a change in customer behavior, and keep the difference unresolved if the source is still unclear. A revised allocation alone is not proof of a performance trend.
