Why app install counts never agree
The web version of this problem, which the platform versus analytics discrepancies guide works through, has two main counters. An app has at least four, and they count different events at different moments under different rules.
| Source | What it counts, as the site's guides describe it | When it counts | User-level detail |
|---|---|---|---|
| Store consoles (App Store Connect, Play Console) | Downloads by the store's own rules, including re-installs and devices that never open the app; installs by source category such as search, browse and referrals | When the store records the download | None; aggregated by category |
| Analytics SDK (for example Firebase feeding Google Analytics) | The first open, collected automatically when the app is launched with the SDK running | On first launch | Yes, within the app's own data |
| Measurement partner | Installs matched to campaigns under the advertiser's windows and rules, with one winner per install after deduplication | On first open, once matched | Yes, as the platforms permit |
| Each ad network | Installs it claims after its own ads, under its own windows; on iOS, delayed and aggregated SKAdNetwork postbacks for users who have not allowed tracking | Varies; iOS postbacks arrive after a delay | Varies; aggregated on iOS |
The table rests on statements the site's guides make from the stores', Apple's and Google's documentation: the measurement plan for app installs on what store consoles count, the deep links and app analytics guide on the first open the SDK collects, and the attribution guide on the partner, the networks, the store reports by source category and SKAdNetwork's delayed, volume-dependent postbacksA server-to-server message reporting a conversion, install or in-app event back to the ad network or affiliate platform that should receive credit. as Apple's documentation describes them. The exact metric names and report definitions in each console, and SKAdNetwork's timers and windows, change and are not repeated here; read them in the current App Store, Play Console and SKAdNetwork documentation when you write your definitions.
Two consequences follow. First, the gaps are mostly structural. A person who downloads and never opens the app is in the store's count and nobody else's; an install claimed by three networks is three network installs and one partner install. Second, the useful signal is change: a gap that has sat in the same place for months and suddenly moves means something broke or something changed, and that is worth knowing quickly. Watching for a gap that leaves its agreed range, rather than debating every month's figures, is management by exception applied to measurement.
Before you start
Get read access to every source: both store consoles, the analytics property, the measurement partner's dashboard or raw export, and each network's reporting. Find the time zone each one reports in and whether it can be changed. Find out which date each uses for an install: the download, the first open, the click or the postback's arrival.
Gather the history you will need: several past months of each source, by platform. Get the app's release history with SDK versions, and the measurement change log if the team keeps one; the measurement change log guide explains what it should hold. Name an owner for the monthly table and the investigation, as the measurement plan for app installs asks.
The steps
List every source that reports installs
Write down every place an install number appears: both store consoles, the analytics property's app streams, the measurement partner, every ad network with an app campaign, and any internal report that copies one of them. Note for each who reads it and for which decision. Reports that quietly copy a source should point back to it, so one change does not have to be traced through several copies.
Write what each one counts
For each source, write the definition in one or two sentences, in the terms the site's guides use and checked against the source's current documentation: what event is counted, whether re-installs count, whether devices that never open the app count, which attribution rules and lookback windowsThe time after an ad interaction in which a conversion can still be credited to it. apply, whether figures are observed or modelled, and what date the install is placed on. Where a definition cannot be found, write "unknown" rather than guessing; an unknown is itself a reason a gap cannot be explained.
Align dates, time zones and platforms
Choose one calendar month and one time zone for the table, and record the offset for any source that cannot be changed. Split everything by platform, because iOS and Android are counted under different rules and a problem on one is hidden in a combined total. Compare only complete months after each source's processing has settled; iOS postbacks in particular arrive after a delay, so the latest days of a month are incomplete for a while.
Build the monthly table
One row per source and platform, one column per month, with the gaps computed between the pairs that matter: store downloads against analytics first opens, analytics first opens against partner installs, partner installs against the sum of network claims, and partner paid installs against each network's own claim. Keep partner organic and paid installs on separate rows, because a broken integration often moves installs from one to the other without changing the total.
Mark the gaps you expect and their direction
Write down, before reading the numbers, which way each gap should point. Store downloads should sit above analytics first opens, because some downloads are never opened and re-installs are counted. The sum of network claims should sit above the partner's paid installs, because networks claim overlapping installs that the partner deduplicates (DeduplicationRemoving double counts when one conversion is reported by more than one source.). Partner installs and analytics first opens should be close in direction and trend, though their rules differ. A gap pointing the wrong way, such as network claims below the partner's figure, is a finding in itself.
Set a tolerance from your own history
Take the months when nothing was known to be broken and look at how far each gap moved from month to month. That spread is the normal range for this app, these networks and this setup. Set the tolerance for each gap from it, and write it down with the owner, as the measurement plan for app installs asks. Do not borrow a figure from anywhere else: the size of each gap depends on the app's audience, its campaigns, its share of iOS users and its partner's configuration.
Investigate when a gap moves
When a gap leaves its range, find out where the move sits before guessing why. Split by platform, then by source category or network, then by app version and country. The guide to explaining metric changes gives the method: rule out data changes first, split the metric into parts, separate mix from rate and find the segment that moved. A move confined to one platform and one app version points to the app; a move confined to one network points to that network's integration or traffic.
Check SDK releases and integrations first
The commonest causes are on the team's own side. Check whether a release changed the analytics or partner SDK, its initialisation or its consent handling, which can drop first opens on one platform; whether a network integration was changed, disconnected or had its windows edited; whether a new campaign or partner started sending traffic that behaves differently, including traffic that may be fraudulent, which the app install campaigns guide describes; and whether a store changed how a report is defined, which its documentation or release notes should say. Fix faults, then rerun the table for the affected months. Catching SDK faults before they ship is cheaper, and the tutorial on checking app events on every release tests each release build's events, consent states and partner postbacks on clean devices.
Record findings in the change log
Write every finding, fault and fix into the measurement change log with the date, the platform, the affected months and the effect on each gap. When a gap's normal range genuinely changes, for example after a deliberate change to attribution windows, record the new baseline and the reason, so next month's reader does not investigate it again.
Decide which number each report uses
Write down which source each recurring report uses and say so on the report. The attribution guide's division is a good default: the partner's deduplicated AttributionAssigning conversion credit to touchpoints. for budget decisions and cost per installAverage ad cost of each app install. by source, each network's own figure for optimising inside that network, analytics for in-app behaviour and cohorts, and the store consoles as the sense check on totals. A report that mixes sources without labels is where most arguments about install numbers start. In the app marketing operations playbook, this monthly table feeds the cohort read that decides which install sources keep their budget.
A template for the monthly table, with the expected direction written beside each gap.
| Line | Platform | Month 1 | Month 2 | Month 3 | Expected direction | Tolerance | Owner |
|---|---|---|---|---|---|---|---|
| Store downloads | iOS / Android | Count | Count | Count | Highest of the counts | Not applicable | Analytics lead |
| Analytics first opens | iOS / Android | Count | Count | Count | Below store downloads | Range from history | Analytics lead |
| Partner installs, organic | iOS / Android | Count | Count | Count | Moves with first opens | Range from history | Growth analyst |
| Partner installs, paid | iOS / Android | Count | Count | Count | Below sum of network claims | Range from history | Growth analyst |
| Sum of network claims | iOS / Android | Count | Count | Count | Above partner paid installs | Range from history | Paid lead |
| Gap: store to first opens | iOS / Android | Share | Share | Share | Positive and stable | Range from history | Analytics lead |
| Gap: networks to partner paid | iOS / Android | Share | Share | Share | Positive and stable | Range from history | Paid lead |
A meal-kit app's monthly reconciliation (illustrative)
Moppetine Meal Kits traces a widening gap
Moppetine Meal Kits, an invented meal-kit app, reconciles each month in a shared table split by platform. Over the six months before the problem, Android store downloads ran 14% to 17% above analytics first opens, iOS 11% to 13%, and the three networks' combined claims ran 35% to 42% above the partner's paid installs. In March the Android store-to-first-open gap jumps to 31% while iOS stays at 12%. The partner's Android installs hold steady and the network gap barely moves, so the problem looks like analytics rather than attribution. Splitting by app version shows the drop is confined to version 8.4, released on 4 March. The release notes show an update to the analytics SDK that changed when it initialises, after the consent screen rather than before the first screen, and a bug left it uninitialised for people who dismissed that screen. The fix ships in 8.4.1 on 19 March; the April gap returns to 16%. The analytics lead records the fault, the affected dates and the estimated missing first opens (about 3,900) in the change log, labels March's Android first opens as incomplete in the monthly report, and adds a first-open check to the release list.
The reconciled numbers feed the reports that decide budgets, so the definitions written here should match the ones the measurement plan framework asks every plan to record, and the mobile measurement hub shows where this routine sits beside attribution and deep links. The glossary entry on deduplication is a quick reference for colleagues who ask why the networks' totals are higher. Installs are only the first count in the chain; the tutorial on finding the app activation event defines the event that turns them into activated users.
Common mistakes
- Chasing a perfect match. The gaps are structural. Aim for a known, stable difference and investigate change.
- Combining platforms. iOS and Android are counted under different rules; a fault on one hides in the total.
- Comparing incomplete months. Delayed iOS postbacks and processing lags make the latest days unreliable for a while.
- Mismatched time zones and dates. One source dating by download and another by first open, in different time zones, manufactures gaps at every month boundary.
- Borrowing a tolerance. The normal gap depends on your app and setup; set it from your own history.
- Summing network claims as the total. Networks overlap; the partner deduplicates. Use the partner for budgets.
- Blaming the store first. Releases, SDK changes and integrations are the commonest causes; check them first.
- Fixing without recording. Without the change log, next quarter's analyst rediscovers the same break.
- Unlabelled reports. Every report should say which source its install number comes from.
