Back to all articlesCRM measurement

Set Attribution Windows Without Rewriting Historical Results Silently

Alactic Editorial10 min read

Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.

A CRM report can show paid acquisition gaining credit even when no new opportunities have closed. Extend the attribution lookback window, rerun historical records, and an older paid touch may displace the source previously credited. The underlying business outcome is unchanged; the measurement rule is not.

To prevent silent rewrites, define the eligible-touch interval precisely, give the complete attribution rule an immutable version, and compare old and new results on the same frozen records. Then make an explicit choice: publish a labelled historical restatement, or preserve the original series and show the new version separately. Never splice the two together as though they were calculated identically.

The worked example below provides a filled sensitivity record and a documented decision. All records, dates, amounts and results in the example are hypothetical. The workflow is a recommended operating design, not a claim about controls available in a particular CRM.

1. Specify what the lookback window actually controls

A lookback window determines which touches are eligible for source credit relative to a defined business event. It does not, by itself, determine which eligible touch wins. A first-touch model and a last-touch model can use the same window and still assign different sources.

Before choosing a duration, write down these parts of the rule:

  • Attribution unit: the entity receiving credit, such as an opportunity rather than an entire account.
  • Anchor event: the timestamp from which the window is measured, such as the opportunity's first transition to closed-won.
  • Eligible touch: the recorded activity types allowed to compete for credit.
  • Identity relationship: which contacts and touches belong to the attribution unit at the reporting cutoff.
  • Interval boundaries: whether touches exactly at the start and end are included.
  • Selection and fallback: how a winning source is chosen and what happens when no eligible touch exists.

For example, an opportunity-creation anchor answers a different question from a closed-won anchor. Changing the anchor while changing the duration makes it difficult to tell why credit moved. Treat those as separate experiments.

Keep this definition separate from campaign response windows and communications rules. A CRM source-credit window describes measurement eligibility; it is not permission to contact someone or a rule for when a message may be sent.

2. Choose a duration by testing the reporting question

Start with the question the report must answer. If the report is intended to describe the earliest recorded source preceding a won opportunity, test whether the proposed duration excludes touches that the business considers relevant to that question. Do not choose a longer window simply because it reduces unattributed records.

Use a fixed historical cohort to inspect the age of eligible touches relative to the chosen anchor. Compare candidate windows while keeping source mapping, identity relationships, selection logic and outcome values unchanged. Otherwise, the comparison measures several changes at once.

For each candidate, examine three observable effects:

  • Coverage: how many records have at least one eligible touch?
  • Credit movement: which records change source, and between which sources?
  • Older-touch dependence: how much new credit comes from touches near the extended boundary?

A longer first-touch window can bring older acquisition activity into view, but it can also give distant touches precedence over recent ones. A shorter window can emphasize recent activity while leaving more records unattributed. Neither outcome proves that a channel caused the sale.

Choose the duration that fits the stated reporting purpose and the available history. If older touch records are incomplete, label that limitation rather than interpreting their absence as evidence that no interaction occurred.

3. Version the complete rule, not just the day count

A label such as “90-day attribution” leaves too much unspecified. Two reports can share that label while using different anchors, time zones or tie-breaking rules.

The following is a hypothetical filled rule record. Its day counts and dates are illustrative, not recommended defaults.

Rule family: Won-opportunity first recorded source
Attribution unit: Opportunity ID
Anchor: First recorded transition to closed-won
Time basis: UTC elapsed time; one day = 24 hours
Interval: [anchor minus window, anchor], both ends inclusive
Eligible touches: Recorded acquisition-source events with
  a valid event timestamp and a mapped source
Identity: Opportunity-contact links frozen at dataset cutoff
Winner: Earliest eligible touch
Tie-break: Lowest immutable event ID, ascending
Fallback: Unattributed; no carry-forward of old credit
Value: Opportunity amount frozen at dataset cutoff

Version W30-v1: Window = 30 elapsed days
Version W90-v2: Window = 90 elapsed days
Other rule components: Unchanged
Dataset snapshot: SNAP-2026-04-03-A
Snapshot cutoff: 2026-04-03 12:00:00 UTC
Proposed effective date: 2026-04-06 00:00:00 UTC

Store the rule version with the calculated result, not only in a dashboard description. Retain the anchor, credited touch ID, credited source, calculation time and dataset snapshot reference. Where practical, retain a reason code such as “no eligible touch” or “earliest eligible event selected.”

A model version and a dataset snapshot solve different problems. The version explains the logic; the snapshot explains the evidence available to that calculation. A later-arriving touch can change a result under the same rule, so preserving only the rule version is insufficient.

4. Run a filled sensitivity test at the boundaries

Fictional worked example: the following cohort contains eight hypothetical won opportunities, each worth $1,000. Each touch age is measured backward from that opportunity's anchor. All listed touches satisfy the eligibility rules unless explicitly excluded by time. The same frozen records are calculated under both versions.

The plain-text sensitivity table includes exact boundaries, a just-outside case, a post-anchor event and a tie. “None” means unattributed, not direct traffic.

Hypothetical sensitivity table
ID | Recorded touches relative to anchor
   | W30-v1 credit | W90-v2 credit | Explanation
---|-----------------------------------------------
A  | Paid: 60 days before; Organic: 10 days before
   | Organic | Paid | Older paid touch becomes eligible
B  | Referral: exactly 30 days before
   | Referral | Referral | Lower boundary is inclusive
C  | Paid: 30 days + 1 second before;
   | Organic: 2 days before
   | Organic | Paid | Paid is just outside W30-v1
D  | Referral: exactly 90 days before
   | None | Referral | Included at W90-v2 boundary
E  | Paid: 90 days + 1 second before
   | None | None | Outside both windows
F  | Email: exactly at anchor
   | Email | Email | Upper boundary is inclusive
G  | Paid: 1 second after anchor
   | None | None | Future touch is excluded
H  | Paid event 100: 20 days before;
   | Organic event 101: same timestamp
   | Paid | Paid | Lower immutable event ID wins

Hypothetical aggregate reconciliation
Source       | W30-v1 count/value | W90-v2 count/value
Paid         | 1 / $1,000         | 3 / $3,000
Organic      | 2 / $2,000         | 0 / $0
Referral     | 1 / $1,000         | 2 / $2,000
Email        | 1 / $1,000         | 1 / $1,000
Unattributed | 3 / $3,000         | 2 / $2,000
Total        | 8 / $8,000         | 8 / $8,000

Three of the eight hypothetical opportunities change source. Two move from organic to paid; one moves from unattributed to referral. Total won value remains $8,000. Paid's additional $2,000 is a measurement reallocation, not additional revenue.

This distinction matters when someone presents the new paid total as evidence of improved performance. In this test, there was no performance change to measure: the cohort, outcomes and recorded touches were held constant.

Run both a boundary test like this and a representative historical-cohort comparison. Boundary tests reveal implementation errors; a broader comparison reveals how much the rule change matters to actual reporting.

5. Decide whether to restate or keep separate series

Restatement is useful when readers need a comparable historical trend under one rule and the underlying records can be recalculated consistently. Separate series are safer when older data is incomplete, previous results must remain reproducible, or the new reporting question differs from the old one.

You can preserve the originally published series while also publishing a restated comparison. The essential requirement is that readers can tell which is which.

Here is the hypothetical decision attached to the sensitivity test:

Decision record: ATTR-2026-04-01
Status: Approved in this fictional example
Decision owner: Revenue operations lead
Reviewers: Marketing operations and finance reporting

Selected rule: W90-v2
Purpose: Include earlier recorded acquisition touches in
  the won-opportunity first-source report

Historical treatment:
- Preserve all previously published W30-v1 reports.
- Publish a separate W90-v2 comparison for January-March
  2026 using snapshot SNAP-2026-04-03-A.
- Label it: Restated under W90-v2; not originally reported.
- Do not overwrite historical W30-v1 source-credit fields.
- Do not restate earlier periods because comparable
  frozen touch history is unavailable in this example.

Forward treatment:
- Use W90-v2 for anchors on or after 6 April 2026 UTC.
- Label any report spanning the transition by version.
- Do not present the mixed-version series as like-for-like.

Required reconciliation:
- Same opportunity IDs and total value in paired runs.
- Explain every source change through credited touch IDs.
- Publish source movement and unattributed totals.

The decision accepts a break in the official series rather than hiding it. For a trend comparison, readers use the separately labelled W90-v2 history and W90-v2 forward results where comparable data exists. The original W30-v1 publications remain available for audit and explanation.

6. Prevent data corrections from masquerading as rule changes

Window changes are not the only reason historical credit moves. A late-imported touch, corrected timestamp, merged contact or remapped source can alter the winner without any change to the lookback duration.

Separate a rule comparison from a data refresh. First run both versions against the same snapshot. If you also need refreshed data, calculate that as a second comparison and explain its effect separately.

  • Late-arriving event: retain both the event time and its arrival time. Decide whether closed reporting periods accept that evidence, and label any resulting revision.
  • Missing or invalid timestamp: flag the record for resolution. Do not silently substitute import time for interaction time.
  • Identity change: identify whether a newly linked contact introduced the winning touch. That is a relationship-data change, not a window-only effect.
  • Source remapping: version the mapping or retain its snapshot. Renaming and regrouping channels can otherwise obscure the movement.
  • Reopened opportunity: apply the stated anchor rule consistently. A new close date should not silently replace the first closed-won anchor in this example.

If a result cannot be reproduced because the original inputs were overwritten, disclose that limitation. Reconstructing an approximation and calling it the original result creates false confidence.

7. Release with reconciliation and a recovery path

Before changing shared reporting, save the old outputs and run the proposed version in parallel. Give reviewers a source-to-source movement summary, the changed opportunity list and the boundary-test results. Large aggregate totals can conceal individual errors, so each moved record needs an explainable winning touch.

Stop the release if the paired runs contain different opportunity populations, different frozen outcome values, unexplained source changes or failed boundary tests. Those failures mean the comparison is not isolating the window change.

If an incorrect rule has already been published, restore the prior labelled output where possible, identify the affected reports and issue a correction. Keep the failed version in the change history, mark it withdrawn and publish the corrected version under a new identifier. Reusing the same version name would make later reproduction ambiguous.

This process can be implemented through controlled exports and review records if the workspace does not support versioned calculations. Check the workspace and current plan before depending on stored attribution versions, parallel runs or historical snapshots; this article does not assume those capabilities exist.

8. Make the next review measurable

For the next proposed change, produce one review packet: the complete rule record, a frozen cohort reference, a filled boundary test, a source-movement reconciliation and a signed historical-treatment decision.

Record the number and share of opportunities that change source, the value moved between sources, the unattributed total under each version and the count of unexplained changes. Require zero unexplained changes before release; business reviewers can then decide whether the explained reallocations fit the reporting purpose.

The goal is not an attribution window that never changes. It is a reporting history in which every change has a named rule, reproducible evidence and an explicit decision about what earlier results now mean.

CRM attributionMeasurement governanceLookback windowsReporting

Put the next conversation to work.

Explore Alactic Communications for WhatsApp conversations, campaigns and AI-assisted workflows. Choose the product and plan that fit your team.

Explore Communications