Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Mobile & App Marketing › Mobile Measurement

App Attribution: MMPs and SKAdNetwork

How app attribution works: what a mobile measurement partner does, install and event postbacks, SKAdNetwork and privacy limits, and modelled results.
Guideadvanced10 min readUpdated

What an MMP does

A mobile measurement partner is a neutral service that sits between an app and the ad networks it buys from. Its SDK runs inside the app, records the install and later events, and matches each install to the ad interaction that most plausibly caused it, under rules the advertiser sets. It then tells each network which installs and events it has been credited with, so that the networks can optimise, and gives the advertiser one deduplicated view across all of them.

The need comes from how networks report. Each network sees only its own ads and claims every install that follows one of them within its window. Run three networks and an install may be claimed three times. The MMP decides which claim wins, applies the same lookback windows and rules to all of them, and becomes the system of record.

FunctionWhat the MMP doesWhy it matters
Install matchingMatches the first open to an ad click or view using the identifiers and signals each platform permitsCredits installs to campaigns where a network alone cannot
DeduplicationPicks one winner when several networks claim the same installStops totals exceeding real installs
Self-attributing networksAsks large networks that attribute internally whether they claim an install, then arbitratesThe largest networks do not share click-level data
Event postbacksForwards chosen in-app events to the credited networkLets campaigns optimise to activation or revenue
SKAdNetwork managementConfigures conversion value schemas and decodes Apple's postbacksMakes iOS campaign data usable
Fraud protectionRejects installs from known fraud patterns before creditingKeeps fake installs out of reports and bills
Cost aggregationPulls spend from network APIsProduces cost per install and per event (CPA) by source
Deep linkingOften provides links and deferred deep linksJoins the link a person tapped to the install that followed

The partner does not replace product analytics. It knows the source of each install and the events sent to it, but it is not built for funnel or feature analysis. Firebase, Google Analytics for apps or a product analytics tool covers that, as the deep links and app analytics guide describes.

Install and event postbacks

A postback is a server-to-server message that tells an ad network something happened: an install it is credited with, or an event performed by a user it acquired. Postbacks are how networks learn. When a campaign is set to optimise towards an in-app event, the network's system decides who to show ads to based on which earlier users performed that event, and it only knows that because the MMP, or the network's own SDK, told it.

Three choices shape postbacks. The first is which events to send. Every network needs the install; event-optimised campaigns also need the target event, and a few supporting events help the network model value. Sending every event the app records is noise. The second is timing. Events sent late, or batched, arrive after the network's optimisation window has moved on, so check that postbacks fire close to real time. The third is consistency. The event the campaign optimises to must be defined identically in the app, the MMP and the network, or the campaign will learn from a different behaviour than the one the business measures.

Postbacks also carry the attribution decision itself. If the MMP credits an install to one network, it tells that network, and the others either receive nothing or receive a notice that they lost. Each network's dashboard will still show its own view, built from its own data and windows. The MMP's numbers are the ones to manage the budget by; the network's numbers are the ones its system uses to learn.

SKAdNetwork and privacy limits

Apple's App Tracking Transparency framework requires an app to ask permission before linking a user's data with data from other companies' apps and websites for advertising. When a person declines, the advertising identifier is not available, and device-level attribution of that install to another company's ad is ruled out. Apple's answer is SKAdNetwork, a framework in which the ad network, the publishing app and the advertised app cooperate, and Apple's own systems send an install postback to the network without identifying the user.

The trade-offs are written into the design, as Apple's developer documentation describes. Postbacks arrive after a timer rather than at once. Detail depends on volume: a campaign with enough installs to protect anonymity receives a finer campaign identifier and conversion value, while a small one receives coarser data or none. Post-install behaviour is compressed into a conversion value that the app updates as the user does things, within limited windows after install, so the advertiser has to decide in advance which few signals matter most. Apple has since introduced AdAttributionKit, which its documentation describes as interoperable with SKAdNetwork and following the same privacy approach.

What that means in practice:

  • Design the conversion value schema deliberately. It is the only post-install signal for most iOS installs. Map it to the activation event, an early revenue indicator, or a combination, and change it rarely so that results stay comparable.
  • Concentrate spend. Many small campaigns each fall below the volume needed for detailed postbacks. Fewer, larger campaigns return more usable data.
  • Read iOS results as delayed and partial. Daily optimisation on iOS data is chasing noise; judge over weeks. The entertainment and gaming marketing guide shows how game studios pair these reports with cohort retention around short launch windows.
  • Do not fingerprint. Apple's policies prohibit combining device signals to identify users without permission. Methods that depend on it are not a workaround but a breach of platform rules.

The wider principles, including consent records and what can be shared with whom, are covered in the analytics section's guide to privacy and consent in measurement.

Android install referrer and platform data

Android has historically allowed more precise attribution. The Google Play Install Referrer, documented in the Play developer help, passes the referrer string from the store link (including campaign parameters) to the app on first open, along with timestamps for the click and the install. An MMP or analytics SDK reads it and matches the install to the campaign that sent the person to the store, deterministically and without a device identifier.

The advertising ID on Android remains available unless the user opts out or deletes it in settings, which supports device-level matching for view-through and for networks that rely on it. Google has proposed further privacy changes to advertising on Android, and their scope and timing have changed more than once; treat any plan that depends on the current state as provisional and check the Play and Android documentation before relying on it.

Platform data also comes from the stores themselves. Play Console and App Store Connect both report installs by source category, such as search, browse and referrals, without user-level detail. These are useful as a sense check: if the MMP and the networks together claim far more installs than the store reports in total, something is being double-counted or invented. The tutorial on reconciling app install numbers turns that sense check into a monthly table with an expected direction and a tolerance for each gap.

Modelled results and incrementality

With iOS data aggregated and some users unreachable, part of every app's paid results is modelled. Networks model conversions they cannot observe; MMPs offer aggregated or predictive reporting that estimates outcomes from partial signals. These are estimates, sometimes good ones, and should be labelled as such in reporting so that nobody treats a modelled figure as a count.

Attribution of any kind, modelled or observed, answers a narrower question than most budgets need. It says which ad preceded an install, not whether the install would have happened anyway. Organic installs, brand searches and word of mouth regularly get credited to whichever ad came last. The question that matters for spend is incremental: how many installs and how much value does this channel add?

Two methods answer it. Incrementality tests compare a group exposed to a campaign with a comparable group that is not, by geography, by time or by randomised audience holdout where the network supports it. Marketing mix modelling (MMM) estimates each channel's contribution from aggregate spend and outcome data over time, without user-level tracking, which makes it naturally suited to privacy limits. The analytics section's marketing mix modelling basics explains the method, and the attribution models framework shows how rules-based credit, modelling and experiments fit together. The guide to choosing an attribution approach helps decide which to rely on for which decision.

Illustrative example — not a real result

Tallyhop Games tests whether its brand search adds installs

Tallyhop Games, an invented puzzle studio, sees its MMP credit 22% of monthly installs to a branded Apple Search Ads campaign costing £18,000 a month at £0.95 per install. It suspects many of those people would have found the game in store search anyway. It pauses the brand campaign in four of its eight English-speaking markets for four weeks and keeps it running in the rest. Total installs in the paused markets fall by 6% against the control markets, not the 22% attribution implied. Tallyhop concludes that most branded installs were organic, cuts the brand budget to £6,000 to defend against competitor bidding, and moves the rest to a prospecting campaign whose incrementality it tests next.

Choosing and configuring an MMP

Choose a partner by the networks it integrates with, the quality of its SKAdNetwork and fraud tooling, how it exports raw data to the business's own warehouse, its privacy commitments and its pricing model. The networks that matter most to the app should be fully integrated, including cost data and event postbacks. Ask how it handles self-attributing networks, how it reports modelled iOS data, and what it costs as install volume grows.

Configuration is where most attribution problems are created. Work through these before spending:

SettingDecisionCommon error
Attribution windowsClick and view lookback windows per network, consistent with the business's modelLeaving each network's default, so sources are judged under different rules
Event listThe activation event, a value event and a few supporting eventsSending everything, or nothing beyond install
Conversion value schemaWhich signals iOS postbacks carryMapping it to revenue alone when revenue arrives too late
Network integrationsPostbacks enabled, cost data connectedIntegrations switched on without event mapping
Fraud rulesRejection rules active before spendTurning protection on after the first fraudulent month
Raw data exportFeed to the warehouse with install and event dataRelying only on dashboards that cannot be joined to revenue
ConsentSDK initialised in line with the app's consent flow and disclosuresSDK collecting before consent where consent is required

Test the setup end to end with a test device and a test campaign link before any real budget runs: install, confirm the attribution, trigger the target event and confirm the postback reaches the network. The app install campaign launch checklist repeats that test before every new campaign, alongside the store, deep link and payback checks.

Before configuring a partner, the measurement plan for app installs decides which tool owns installs, behaviour and spend, and how conversion values encode early value.

How these attribution limits shape a Google App campaign, from the choice of bidding event to separate iOS and Android campaigns, is covered in the Google Ads setup for app installs.

The app installs objective hub explains why one measurement partner should own attribution across every install channel the plan funds.

For the paid social side of an install plan, paid social for app installs turns conversion value planning and deduplication into channel tactics and a budget limited by cohort payback.

When late or missing postbacks may be behind climbing campaign costs, the diagnosis of a rising app cost per install checks the signal reaching each network alongside creative, the store page, the auction and scale.

Common mistakes

  • Trusting network dashboards as the total. Each network claims its own installs; the totals overlap. Manage budget from the MMP's deduplicated view.
  • Default windows everywhere. Different lookback windows make networks incomparable. Set them deliberately and consistently.
  • An afterthought conversion value schema. On iOS it is the only post-install signal most installs carry. Design it around activation and early value before launch.
  • Spreading iOS spend thin. Small campaigns fall below the volume needed for detailed postbacks. Consolidate.
  • Optimising daily on delayed data. SKAdNetwork postbacks arrive late by design. Judge iOS campaigns over weeks.
  • Treating attribution as incrementality. Credit is not contribution. Test channels with holdouts or a mix model before scaling or cutting them.
  • Fingerprinting to recover lost data. It breaches Apple's policies and the trust of users who declined tracking.
  • Postbacks that never reach the network. Event-optimised campaigns starve without them. Test every integration end to end.

Frequently asked questions

Do I need a mobile measurement partner?

If the app runs paid campaigns on more than one network, almost certainly. Each network attributes installs to itself under its own rules, so without a neutral party the totals overlap and cannot be compared. An app on a single network with modest spend can start with that network's reporting and a free analytics SDK, and add a partner when a second network or meaningful budget arrives.

Why are SKAdNetwork results delayed and incomplete?

Because they are designed to be. Apple's framework sends install postbacks to the ad network after a timer, withholds or coarsens detail when the campaign's volume is too low to protect anonymity, and limits post-install information to a small conversion value. The delay and coarseness are what stop the data being joined to an individual. Plan campaigns and reporting around them rather than trying to work around them.

Can modelled or probabilistic attribution replace device-level data on iOS?

Not as a like-for-like replacement. Apple's policies prohibit fingerprinting to identify devices without permission, so methods that rely on it are not an option. Aggregate modelling, such as marketing mix models and incrementality tests, answers a different question (how much a channel adds overall) and is the honest complement to SKAdNetwork's limited campaign-level data.

Terms used on this page

Frameworks

Where this page is used

Also related

Sources

  1. SKAdNetwork — Apple Developer
  2. Google Play Console Help — Google
  3. Firebase documentation — Google

Related