Key takeaway
Confirm both numbers answer the same question. Align five definitions: time window and cutoff, deduplication and merged records, which statuses count, attribution, and source system. A metric dictionary must let someone else recompute the number, with refresh time and data cutoff stated.

Take this illustrative scenario. At the monthly management meeting, the sales lead opens a dashboard: 120 orders closed last month. The finance lead opens a report: 108 orders corresponding to recognized revenue. Neither believes they calculated wrong. The boss asks: where did the missing 12 go?
Those 12 can have five entirely different explanations: different time windows, duplicate records not removed, statuses counted on one side but not the other, different ownership rules, or different data sources. Align the definitions first; only then argue about who is wrong.
Confirm both numbers answer the same question
"Order count" can mean at least four things: orders created, orders paid, orders shipped, or orders with recognized revenue. Those numbers differ by nature, and comparing them directly is meaningless.
So the first step is not SQL. Ask each side to state in one sentence what question the number answers, what time range it covers, and what moment it is current as of. If the sentences differ, the gap may not be an error; two metrics may simply share one name.
Names should reflect the question. "Orders closed" on the dashboard and "orders with recognized revenue" in the finance report is clearer than calling both "order count."
Align five definitions
Time. Is it counted by order time, payment time, shipment time, or financial recognition time? Is the range a calendar month, a rolling 30 days, or up to midnight yesterday? Cross-time-zone business must state the zone. If one team uses a calendar month and the other uses data through yesterday, an early-month comparison will always differ.
Deduplication. Is a row an order or an order line? Do repeated submissions, split orders, and merged customers count once or several times? Does a re-order after a refund count as new or as the same? Write the dedupe key: order number, payment transaction ID, or customer plus time.
Status. Do drafts count? Cancelled orders? Refunded orders? Which status does a partial shipment take? Status also changes over time: an order pending shipment last month is completed this month, so two sides taking numbers at different moments naturally disagree.
Ownership. Is the sale attributed to the person who entered it, the person who collected payment, or the branch? If the customer later moves to another salesperson, do historical orders move too? Do channel orders count as direct or channel? Change the rule, and the same order changes owner overnight.
Source. Does the number come from the operational database, the general ledger, a data warehouse, or a manual spreadsheet? The operational system may record by action time while finance records by recognition time; a warehouse sync may lag, leaving the dashboard behind. Different sources require deciding which one is authoritative for which question.
Write the metric dictionary so someone else can recompute it
A usable metric dictionary has at least nine parts:
- Name: a business-readable name, not an internal abbreviation.
- One-sentence definition: what question the number answers.
- Formula: count or sum, numerator and denominator, stated exactly.
- Time field: which timestamp filters the range, and how the range is chosen.
- Included statuses: what counts and what is excluded.
- Dedupe key: what makes two rows the same event.
- Source: which table or system, and whether it is synchronized.
- Refresh and owner: how often it updates and whom to ask when it breaks.
- Known limits: for example, "excludes offline manual orders" or "may change after financial close."
Then test it: can someone who did not write the definition compute the same number from the dictionary alone? If not, it is not clear enough.
Refresh time and data cutoff
The dashboard should state "Data through yesterday 24:00" or "Updated today 09:15." A real-time number without a cutoff invites the assumption that it includes everything up to this second.
A live operational view and a financial close legitimately differ. Operations needs timely trends; finance needs accurate recognition. They do not have to be forced into one number, but each must state its purpose and cutoff.
Late-arriving data also needs stating. If the warehouse backfills yesterday's shipment today, yesterday's dashboard number will change. If the dashboard overwrites history, tell users that history can be adjusted by backfills, so a screenshot from yesterday and today's page do not appear to contradict each other without explanation.
Let operational and financial figures coexist
When both numbers are useful, show them side by side with their definitions rather than overwriting one with the other:
- Operations dashboard: paid orders, by order time, data through today 09:00, used for follow-up cadence.
- Finance report: orders with recognized revenue, by recognition time, data through last month's close, used for settlement.
If management needs one shared number, designate a "management definition," write it in the dictionary, and state how it converts to the others. Do not average two undefined numbers and call the result the truth.
A sample metric dictionary
Field | Example content |
|---|---|
Name | Paid orders |
One-sentence definition | Orders that completed payment in the period and are not fully refunded |
Formula | Count of distinct order numbers |
Time field | Payment success time |
Included statuses | Paid, partially refunded; excludes unpaid, cancelled, fully refunded |
Dedupe key | Order number |
Source | Operational payment table, synced hourly |
Refresh and owner | Hourly; operations data team |
Known limits | Excludes offline manual orders; refund status lags |
Keep the dictionary link beside the dashboard title, so anyone looking at a number can open its definition, cutoff, and owner. When two numbers disagree, open the dictionary before opening the database.
Take the most contested metric and fill in this table. Doing so, the two departments often discover that of the 12 missing orders, five differ by time window, four are duplicate submissions, and three differ by refund status. Only the remainder is a genuine data problem to investigate.
Which modules a dashboard should hold, and whose next step it serves, is covered in the dashboard design discussed in the system preparation guide; this article covers numeric definitions only. The fields and sources behind a metric belong in the early scope of business system development. How weekly-report automation makes statistical definitions explicit is covered in the weekly report automation chain.