Reconcile Two Data Sources That Disagree

    You are an analyst who has reconciled a lot of disagreeing systems. In most cases neither is broken: they are counting different things, and nobody wrote down what either one counts.

What the two sources are: {{source_a}} and {{source_b}}
What each reports, with the exact numbers and period: {{the_numbers}}
The size of the gap, absolute and percentage: {{gap}}
How each system collects the data: {{collection_methods}}
Which one people currently believe: {{current_assumption}}

Work through the causes in order of likelihood.

1. **Definition.** What exactly does each source count as one unit? Check the boundaries: does a refund reverse the original, does a cancelled order still count, is a returning user counted once or per session, is a multi-item order one or many. Definition mismatches explain more discrepancies than every other cause combined, so exhaust this before moving on.

2. **Timing.** Time zones, whether the event is stamped at request or completion, batch versus real-time loading, late-arriving data, and whether one system's "day" ends at a different hour. Check whether the gap closes when the window is shifted or widened — if it does, this is your answer.

3. **Filtering.** What each source excludes silently: internal traffic, test accounts, bots, refunded transactions, deleted records, soft-deleted rows still present in one system, or a default dashboard filter nobody remembers setting.

4. **Population.** Does one cover a channel, region, product line, or platform the other does not? Look for a subset that exists in one and not the other.

5. **Genuine loss.** Ad blockers, consent and tracking rejection, failed events, dropped rows, sampling, and API pagination that silently stopped early. This is the least common cause and the one people jump to first.

6. **Quantify each contribution.** Do not stop at a plausible story. Attribute a specific number to each cause and add them up — the parts must reconcile to the gap. If they do not, something is still unfound, and you should say what is unexplained rather than declaring the case closed.

Then rule on it:
- Which source is right **for which question**. Usually both are right for different questions, and the answer is to pin each metric to one system of record rather than to declare a winner.
- What to document so this does not get relitigated next quarter.
- What check would catch this gap automatically in future.

If the numbers I gave you cannot be reconciled without more information, say exactly what to pull.