A retail result needs a traceable path from source to decision. A dashboard can display a precise number even when its period, item population or denominator does not match the question.
Start by defining what the number means and what evidence would make it usable. The workflow below is my recommended quality checklist for supplier and advisor teams; it is not an official retailer specification.
Choose the right source for the question
Use the source that represents the business concept being measured. Retailer POS can describe shopper sales in its stated scope. A supplier's own ledger answers different questions about invoiced revenue, costs or contribution. An advertising report assigns credit under its measurement rules. Those numbers should not be forced to match simply because each is labeled “sales.”
In Scintilla, distinguish sales and inventory analysis from shopper behavior analysis. Confirm the client's Basic or Charter plan, permissions, report definition and available fields before designing a pull. A report's presence in documentation does not establish access for every supplier or advisor.
Keep Walmart and Sam's Club data requirements explicit. Walmart announced Scintilla's planned expansion to Sam's Club in 2027; that announcement does not establish that a Sam's Club report is available in a current Walmart account. Verify the present source and entitlement for the work. Walmart's August 20, 2026 announcement.
Choose a primary source for each result, and reconcile meaningful differences with other sources before interpretation. If a material discrepancy appears during a test, investigate it then; do not preserve a known error merely to keep the original source choice unchanged.
Write a metric contract
A short metric contract should record:
- The business question, accountable owner and metric name.
- Retailer, supplier account, item population, channel and geographic scope.
- Numerator, denominator, units and aggregation rule.
- Date basis, fiscal or calendar period, time zone and comparison mapping.
- Returns, cancellations, exclusions and missing-data handling.
- Source, extraction time, source version and restatement policy.
Define the actual record grain. “Store, item, date” may identify one daily aggregate; it is not a safe universal key for transactions, channel splits, order lines or inventory snapshots. Deduplicating on an incomplete key can delete legitimate activity.
For rates, aggregate the eligible numerator and denominator according to the metric's definition. Do not average item percentages unless an equal-weighted item average is the intended measure. Distinguish inventory at a point in time from sales over a period, and keep eaches, cases and variable-weight quantities separate unless an explicit conversion is valid.
Preserve the clocks and mappings that matter
Keep the source's business-date rule. A transaction date, shipment date, inventory observation date and advertising touchpoint date answer different questions. Processing and extraction times also matter because they show what was available at the time of the decision.
Retain item and store mappings, including effective dates and any succession or hierarchy changes. Save the evidence used for the original read. If the source later restates history, record the change and rerun the relevant comparison on a consistent basis. Neither silently mixing extracts nor freezing an acknowledged error produces a reliable trend.
Apply six practical checks
| Check | Question to answer |
|---|---|
| Freshness | Has the expected source period arrived, and is the comparison equally mature? |
| Completeness | Is the eligible population represented? Are absent rows missing, suppressed, out of scope or valid zero activity? |
| Uniqueness | Are records unique at the documented grain, including channel and other necessary dimensions? |
| Validity | Do units, signs, dates and values comply with the field's meaning? Net sales or adjustments may legitimately be negative. |
| Reconciliation | Do totals and breakdowns reconcile within the same scope? Are overlaps and unshown remainders explicit? |
| Change detection | Has a schema, definition, mapping or source population changed in a way that affects interpretation? |
Set tolerances based on materiality and source behavior. An arbitrary “small difference” is not a reconciliation rule. Missingness concentrated in one test group can matter even when overall coverage looks high.
Check the experiment as well as the pipeline
Compare recorded assignment with the intended design, then verify actual execution separately. Keep an exception log for price, promotion, assortment, media and field changes. Correct defects through the authorized owners and preserve the history of what happened.
An A/A run, where both groups receive the same treatment, can help test measurement behavior. A chance difference can still occur, and a quiet A/A does not certify all future comparisons. Investigate unexpected assignment patterns and result distributions before making the business claim.
The retail test planning guide covers assignment and uncertainty. Reliable input data supports that work; it does not by itself establish causation.
Use AI after the evidence is defined
Compute and check metrics with reproducible queries or calculations. Give an AI assistant a reviewed, appropriately shareable summary and ask it to explain observations, flag missing context and draft questions. Review every number and source reference before using the output in a meeting.
Retailer data and client material have their own use restrictions. Access to an export does not automatically authorize uploading it to an external AI service or using it for model training. Keep each client's information within the approved workflow.
Retail Reason Intelligence can help clarify retailer terminology, operating context and questions to investigate.
When the next question needs work inside Scintilla Charter, explore direct advisory support for report design, population checks, test planning, and interpretation across the relevant applications.