Reconcile the numbers before asking AI to explain them
Two dashboards can disagree for ordinary reasons. Establish their definitions before turning the difference into a business story.
Norla Editorial5 min
A weekly report can contain a persuasive explanation of a change that exists only because two sources use different definitions. One export may count conversion events while another counts orders. One may use event time while another uses processing time. Before asking a model to explain performance, establish whether the quantities can be compared.
Document the meaning of one row, the filters, the time zone and the reporting period for each source. Then inspect currency, refunds, taxes and duplicate treatment where those fields matter. Keep these choices in a short data contract so the next operator does not repeat the same investigation from memory.
Reconciliation should preserve unmatched records and unresolved differences. A joined table that quietly drops missing IDs can look precise while removing the very cases that explain the gap. Report the match rate and the nature of the remaining records, and avoid turning every residual difference into an attribution claim.
Once the sources are aligned, a narrative assistant can help organize observations and questions. Ask it to separate verified changes from hypotheses and to link each quantitative statement to the relevant table. The useful output is a reviewable explanation with clear limits, not a smooth paragraph that conceals inconsistent inputs.
Put it into practice
Record the row meaning and time basis of each source.
Check filters, currency and refund treatment.
Retain unmatched rows in a separate review table.
Label business explanations as hypotheses until supported.
Specific questions, practical answers, and the next detail to check. Prepared by Norla Editorial.
Q
Question 01
Two exports agree on total order value but disagree on the number of orders. The reporting deadline is close. What is the smallest useful investigation before preparing the management summary?
A
Norla Editorial · Answer
Check the unit of each row, duplicate identifiers, aggregation rules, and treatment of canceled records. Agreement in value does not establish that the records represent the same population. Prepare a short exception table and explain whether the unresolved count difference could affect the specific decision the summary is intended to support.
↳
Follow-up question
If the decision concerns total value rather than order count, can the team publish the value comparison while leaving the count discrepancy unresolved?
A
Norla Editorial · Clarification
It may prepare a qualified comparison if the owner accepts the remaining uncertainty and there is evidence that the discrepancy does not undermine that measure. State the unresolved issue beside the result. Do not present the whole dataset as reconciled merely because one decision-relevant total appears consistent.
Q
Question 02
A reporting spreadsheet uses local calendar dates, while a second export includes timestamps from another time zone. How should an analyst define the comparison before asking for an explanation of daily differences?
A
Norla Editorial · Answer
Choose the reporting time zone and the meaning of each date field, then document how records are assigned to a reporting day. Preserve the original timestamps where authorized. Rebuild the daily comparison under that explicit rule before interpreting a shift as customer behavior or campaign performance.
↳
Follow-up question
What if one source contains only a date and does not provide enough information to translate each record into the chosen reporting time zone?
A
Norla Editorial · Clarification
Do not invent missing timing information. Identify the boundary uncertainty and consider a broader comparison window if that still serves the decision. Keep any remaining mismatch visible. The analyst should explain what the available date field supports rather than implying that a precise conversion has been performed.
YOUR SIDE OF THE DISCUSSION
Add your perspective.
Your own notes stay private on this device. They are not sent to other members.
Background discussion & source notes
NORLA EDITORIAL / DISCUSSION DESK
Let’s take the question further.
Practical follow-ups, open questions and considered answers from the Norla editorial desk.
5 official discussion notes
Norla Editorial@norla.editorial · Note 01
Start with a shared comparison unit
When two reports disagree, first ask whether they are counting the same thing. A campaign report may count attributed conversions while an order export records orders after cancellations. A narrative about weak creative performance cannot resolve that definitional mismatch. Choose a comparison unit, reporting window, currency treatment, and rule for adjustments before generating explanations. Preserve each system's original total and document the transformation used for comparison. The first useful artifact is often a reconciliation table with explicit unmatched groups, not a confident executive summary that makes unequal definitions sound like a business trend.
Consider a Monday review comparing yesterday's analytics data with a finance export prepared before the weekend. Recent data may still change as processing completes; Google Analytics documents data freshness and processing considerations. The two files also have different extraction times, which creates another source of disagreement. Record both timestamps and label the comparison period before escalating a discrepancy. For this exercise, separate timing differences from record mismatches and definition differences. That classification gives the team a concrete investigation queue without asserting that one platform must be wrong or that the latest visible figure is final.
Record-level matching can expose duplicates and unmatched transactions when stable identifiers are available and their use is authorized. Aggregate comparison can be appropriate when only summarized exports are permitted, but it cannot answer every question about individual records. Do not force a row-level story from totals alone. State which method the available data supports and what remains unresolved. A combined approach may reconcile daily totals first and inspect a small authorized exception set afterward. The trade-off is between diagnostic detail and the access, preparation, and privacy requirements necessary to obtain that detail responsibly.
Two totals can agree while individual records have been duplicated and others omitted. A common shortcut checks only the final sum after merging exports. Instead, inspect the intended relationship between tables, the number of records before and after the merge, and whether one identifier maps to multiple rows. Keep a small exception table that shows why records were excluded or grouped. A balancing adjustment may be useful for investigation, but it should never silently replace an unexplained difference. The goal is a reproducible comparison, not a spreadsheet whose final line happens to look reassuring.
Bring three pages to the review: definitions, reconciliation, and unresolved questions. For each material difference, list the observed amount or count, the records or group involved, the proposed explanation, and the check needed to confirm it. Keep a hypothesis visibly separate from a finding. Ask the business owner which differences could change the decision under discussion; investigate those before polishing the narrative. The next useful debate is whether the remaining uncertainty is acceptable for that decision. It is not necessary to resolve every historical mismatch before making a carefully bounded operational choice.
NORLA EDITORIAL / FIELD NOTES
Continue exploring.
Working methods, decisions to document, and useful questions to take into your next project.