Measurement
Part of Measuring B2B social selling with four evidence levels kept apart
Correcting B2B social selling measurement mistakes: preserve, repair, restate
Correct a B2B social selling measurement mistake in England: preserve the release, record the defect, repair the data path and issue a dated restatement.
A published B2B social selling figure that turns out to be wrong needs a visible correction, not a quiet edit. Preserve the affected release, record the defect, repair the definition or data path, and issue a dated restatement. That sequence keeps the evidence needed to prevent the same error next month.
What to take away
- Correct a published figure in four movespreserve, record, repair and restate.
- Never change a chart quietly, because that erases the evidence of the error.
- Classify the defect first, since each type of mistake has a different repair.
- Reconcile source rows, loaded rows, duplicates, unmatched records and corrections for each refresh.
- Date the restatement and name the cause, so the same error is easier to prevent.
Preserve first, then repair
Preserve the affected release before touching anything. The earlier version is the evidence of what was published, and the repair can only be judged against it. Quietly changing a chart erases the evidence needed to prevent the same error next month.
- Preserve the affected release exactly as it was published.
- Record the defectwhich figure, which definition or data path, and what went wrong.
- Repair the definition or data path, not just the displayed number.
- Issue a dated restatement that names the old figure, the new figure and the cause.
The repair belongs at the source. A corrected chart on top of a faulty extract, transformation or match will fail at the next refresh. The stored extract, transformation and review are the evidence behind a dashboard, which is only a decision interface.
Classify the defect before choosing a repair
Six types of defect call for different repairs. Name yours before repairing it, because a restatement that fixes the wrong cause will itself need restating, and each repair below changes what the corrected report must say.
| Defect | What the report did | Repair before restating |
|---|---|---|
| Activity as outcome | Showed messages, profile views or saved leads as effectiveness | Keep the activity count separate and pair it with a later outcome with its own defined population |
| Rate without denominator | Gave a percentage with an unclear base | Write numerator, denominator, cohort entry date, 30-day window and exclusions beside it |
| Attribution as cause | Presented CRM credit as impact | Label it attributed and publish the model, eligible touches and lookback period |
| Unlike cohorts | Read a small-account quarter against a large-enterprise annual average | Freeze eligibility, unit, geography and window, then choose the reference set |
| Dashboard data failure | Refreshed a bad match perfectly | Reconcile source rows, loaded rows, duplicates and unmatched records |
| Worker data reuse | Used adoption counts in performance decisions | State the purpose, aggregate and restrict access |
Restate an attribution claim as attribution
If the defect is attribution presented as impact, restate the claim and not just the number. Attribution assigns credit under a rule. Causal analysis asks what would have happened without the intervention, and a CRM figure cannot answer that question.
HM Treasury's guidance on quality in policy impact evaluation says simply monitoring outcomes shows whether a change occurred, but not whether the intervention caused it. Label the figure "attributed" and publish the model, eligible touches and lookback period, as set out in the comparison of attribution methods.
Missing-path share is eligible opportunities without a complete required touch path divided by all eligible opportunities in the period. Exclude open observation windows and show them separately. Use "incremental" only when a suitable design and its assumptions have received statistical review.
Restate a rate with its full specification
If the defect is a missing denominator, restate the rate with its full specification. A response percentage is uninterpretable when the reader cannot tell whether the base is messages, delivered contacts or unique accounts. Write the base down beside the number.
Give the numerator, denominator, cohort entry date, 30-day response window and exclusions. Keep silence in the denominator and show unresolved delivery status separately. Complete formulas turn a corrected percentage into one that another analyst can reproduce.
For an activity count, the same discipline applies. Approved messages sent is the number released in the calendar month, excluding tests and verified delivery failures. Pair it with a later outcome whose own population is defined, and do not rename usage as effectiveness.
Rebuild a comparison under frozen rules
If the defect is an unlike-cohort comparison, rebuild the reference set under frozen rules. A small-account quarter cannot be read against a large-enterprise annual average without testing comparability, and England and UK totals are not interchangeable.
Freeze eligibility, account unit, sector, size, geography, channel and observation window before selecting the reference set. Publish exclusions and count records that could not be classified. A benchmark range built this way still has weaknesses, and the restatement should say what they are.
Repair the pipeline, not only the page
Automatic refresh can reproduce a bad match perfectly, so repair the pipeline and not only the page. Reconcile source rows, loaded rows, duplicates, unmatched records and corrections for each refresh, and keep the result with the release it checked.
The Government Analysis Function's reproducible analytical pipelines guidance says that where a manual step is absolutely necessary, it must be documented. Treat a manual fix to a metric as such a step and write it down. A dashboard that reconciles every refresh makes the next defect visible earlier.
Stop worker data reuse before restating
If the defect is worker data reuse, stop the reuse before restating anything. Individual login or message counts may be collected for adoption support and later reused in performance decisions, and the restatement should record that the purpose changed.
The ICO's UK worker-monitoring guidance warns about function creep, inaccurate inference and the need for workers to understand and challenge information used about them. State the purpose before collection, aggregate where possible and restrict access. If an individual measure informs appraisal or discipline, obtain qualified data-protection and employment review. A neat seller leaderboard does not repair an unjustified purpose or a misleading event count.
Write the restatement note
A restatement note is short and fixed in form. It names the figure, the old value, the new value, the cause and the repair, and it carries the date on which it was issued.
Restated on [date]: [figure] was published as [old value] in the [original release]. The correct value is [new value]. Cause: [defect]. Repair: [what changed in the definition or data path].
Before you act
- Preserve the affected release before changing anything.
- Record the defect and classify its type.
- Repair the definition or data path, not only the chart.
- Issue a dated restatement naming old figure, new figure and cause.
- Stop any reuse of worker data until its purpose is stated.
Common questions
Why not simply fix the chart?
Quietly changing a chart erases the evidence needed to prevent the same error next month. Preserve the release, record the defect, repair the definition or data path and issue a dated restatement instead.
What should a restatement say?
It should name the figure, the old value, the new value, the cause and the repair, and carry the date it was issued. A reader should be able to compare it with the preserved original release.
How do I restate an attribution claim?
Label the figure attributed and publish the model, eligible touches and lookback period. Use the word incremental only when a suitable design and its assumptions have received statistical review.
