Operations · 2 February 2026
Reading an MMP discrepancy report without panic
The report is not an accusation. It is a list of reasons two ledgers counted different days. Most of those reasons are boring. Boring is good.
Every MMP worth using will eventually show you a discrepancy view: network-reported installs or events on one side, MMP-attributed on the other. New analysts treat a 20% gap like a fire. Veterans look for the timezone first.
The usual suspects, in the order we check them
Timezones. A network closing on UTC and an MMP closing on ICT will disagree every afternoon. If the gap moves with the clock and vanishes on a 48-hour rollup, stop writing incident tickets.
Windows. View-through still on at the network, off at the MMP. Click lookback 7 days versus 1. Reattribution on for one partner and not the other. These are policy mismatches, not “lost data.”
Unpaid and owned. Organic, email, and CRM deep links often land in the MMP and never in the paid network. A discrepancy report that ignores owned traffic will look like the network is inventing installs. Sometimes it is. Sometimes you forgot a calendar blast.
Rejected events. Duplicate purchases, sandbox traffic, events without a customer ID when your MMP rule requires one. Read the rejection log before you phone the account manager.
The finance question
Finance rarely asks why the MMP and Meta disagree by 14%. They ask whether recognised revenue in the general ledger can be tied to a campaign without double counting refunds. If your discrepancy ritual never touches refunds, store chargebacks, or subscription churn, you are reconciling vanity against vanity.
In the lab we force the third column into the workbook: network, MMP, finance. The homework is a paragraph naming which gap you will not close this quarter. That paragraph is the deliverable. The chart is optional.
When panic is warranted
If CTIT clusters under two seconds at scale, if install volume outruns store truth by a factor that survives timezone correction, or if a partner’s self-attributing stack suddenly absorbs organic, then yes — escalate. Bring timestamps, not adjectives.
The Measurement Lab spends a full week on this workbook. The fraud clinic is shorter and meaner if you already know the discrepancy is not a timezone.