The Death of Last-Click in Mobile-App Attribution

TL;DR: Last-click attribution as the operational measurement for mobile app campaigns no longer functions, and the failure modes are now structural rather than methodological. Apple's IDFA opt-out rates above 70% on most apps, the SKAdNetwork postback-loss problem (where a meaningful share of installs never produce an attributable postback at all), the 25-conversion privacy threshold below which Apple returns no data, and the noise added to conversion values have collectively shifted the ground truth from deterministic attribution to incrementality testing. The mobile measurement programs that work in 2024 treat MMP attribution as a directional signal at best and treat geo-holdout incrementality as the basis for budget decisions.

A note on the named companies and sources. Apple's developer documentation on SKAdNetwork (versions 2 through 4), the AppsFlyer State of App Marketing reports, the Adjust mobile measurement guides, the Branch attribution documentation, and the Mobile Measurement Partner industry conference materials appear throughout as available public reference points. Quantitative ranges framed as advisory observation come from anonymized partner app operators in the 100K to 12M monthly active user range across gaming, fintech, and consumer subscription verticals, not from the cited MMPs or platforms.

---

What Broke, and When

The pre-2021 mobile attribution model rested on a small number of deterministic identifiers, primarily the IDFA on iOS and the GAID on Android. The identifiers were available by default to apps and ad networks; they let the attribution layer match an ad click to an install with near-perfect deterministic resolution. The MMPs (AppsFlyer, Adjust, Branch, Kochava, Singular, and a long tail of smaller vendors) built their businesses on this resolution. Last-click attribution on a mobile install was, mechanically, a deterministic lookup: the ad network reported a click, the app reported an install, the MMP matched the IDFA, and the credit went to the network whose click came most recently.

iOS 14.5 in April 2021 introduced App Tracking Transparency, the prompt that asks users whether the app may track them across other apps and websites. The IDFA became available only to apps the user explicitly authorized. The opt-in rates settled, after the early weeks of variance, in the 18 to 32 percent range for most consumer apps, with significant verticals (utility apps, social, gaming) running in the 12 to 24 percent range, and a smaller set of trust-heavy categories (banking, healthcare) seeing opt-in rates above 35 percent. The implication: for roughly 70 to 82 percent of iOS installs, the IDFA was no longer available, and the deterministic attribution chain was broken.

Apple's response to the loss of cross-app tracking was the SKAdNetwork framework, which provides aggregated, privacy-preserving attribution data. The framework has gone through multiple major versions (SKAN 2.0 in 2018, SKAN 3.0 in 2021, SKAN 4.0 in late 2022). The current SKAN 4 version offers richer conversion windows, hierarchical conversion values, and multi-postback support. The framework provides genuine attribution data, but with three structural limitations that have proven operationally significant: postback loss (a share of installs never generate a postback), the privacy threshold (Apple returns null values below 25 conversions per attribution group), and conversion value noise (Apple injects randomness into the returned values to preserve privacy). Each limitation has implications for how attribution data should be interpreted.

The Three Structural SKAdNetwork Limitations

The SKAdNetwork framework's design choices reflect Apple's privacy priorities, and each choice constrains the attribution data in operationally meaningful ways. The three limitations that matter for measurement programs are postback loss, the privacy threshold, and conversion value noise.

Postback loss is the situation where an install occurs but Apple does not deliver a postback to the attributed ad network. The loss arises from several mechanisms: the user uninstalls the app before the postback fires, the device falls below some idle-time threshold Apple uses to batch postbacks, the user's network conditions interrupt the postback delivery, the install attribution window expires, or the install occurs through a path SKAdNetwork does not support (web-to-app, deep link, organic). The observed postback-to-install ratio in partner data ranges from 64 to 89 percent on iOS, with the higher end on apps with longer engagement windows and the lower end on apps with high early-uninstall rates. The 11 to 36 percent of installs that never produce a postback are not "lost" in the sense of being unattributable in aggregate; they show up in the app's own install count. But they are lost in the sense of not being attributable to the specific ad network that drove them, which is the question last-click attribution is supposed to answer.

The privacy threshold is Apple's rule that conversion values are returned as null until at least 25 conversions have occurred in the relevant attribution group (the combination of source app, campaign ID, and conversion window). The threshold is a privacy-preservation mechanism: small-sample data could be used to infer information about individual users, so Apple returns no data until the sample is large enough that aggregation provides cover. The operational consequence is that any campaign producing fewer than 25 conversions in the window returns no conversion-value data, only the install count. For long-tail campaigns, geo-targeted experiments, niche targeting, or small-budget tests, the threshold is binding and the campaign's value performance is invisible.

Conversion value noise is the third limitation. Apple injects randomness into the returned conversion values to provide additional privacy protection, and the noise is meaningful: in SKAN 4, the conversion value can be returned with up to a roughly 12 percent error margin on individual postbacks, with the error averaging out across larger samples but persisting on smaller ones. For an attribution layer trying to compute precise ROAS at the campaign or creative level, the noise interacts with the privacy threshold to produce a measurement floor below which signal-to-noise is too poor to be useful.

The combination of these limitations produces an attribution layer that is genuinely useful at the network and campaign-cluster level for large-budget operators with high-volume traffic, and is operationally useless at the creative, ad-set, or small-campaign level for most operators. The MMPs have layered probabilistic matching, modeled installs, and various proprietary methodologies on top of the SKAN data to recover some of the lost resolution, but the recovery is partial and the residual error is material.

The Probabilistic Attribution Compromise

The MMPs responded to the IDFA collapse partly through probabilistic attribution, the practice of matching ad clicks to installs using non-deterministic signals: IP address, user agent, device model, screen resolution, timestamp, geographic location, and other fingerprintable characteristics. The probabilistic approach was available before iOS 14.5 (it was a backup when IDFA was missing for any reason), but its share of attributed installs grew substantially after the IDFA collapse, and its limitations became more operationally significant.

The probabilistic match accuracy depends on the signal richness available in the click data and the install data. In partner data, the probabilistic-match accuracy on iOS (validated against the small remaining IDFA-opt-in sample as a ground-truth proxy) runs in the 64 to 84 percent range, with the higher accuracy on networks that capture rich click-side signals and the lower accuracy on networks (or networks plus ad-format combinations) where the signal is impoverished. The 16 to 36 percent of probabilistic attributions that are incorrect get distributed across networks roughly in proportion to network volume, which means the larger networks accumulate small over-attribution errors and the smaller networks accumulate small under-attribution errors, but the errors are not zero-sum at the operator level: some installs get attributed to the wrong network, others get attributed to organic when they were paid, others get attributed to paid when they were organic.

Apple's position on probabilistic attribution has been hostile from the beginning. The App Store Review Guidelines prohibit fingerprinting users, and Apple has issued guidance to ad networks and MMPs that probabilistic matching using fingerprintable signals is not consistent with the spirit of the privacy framework. The enforcement has been uneven (the MMPs continue to operate probabilistic matching with various justifications about which signals are or are not fingerprinting), but the operational risk for an operator relying heavily on probabilistic attribution is non-trivial: if Apple tightens enforcement, the matching layer could degrade meaningfully overnight.

The Android equivalent (GAID, Google Advertising ID) was less affected by the iOS 14.5 transition but has been on a parallel privacy trajectory. Google announced in 2022 the Privacy Sandbox initiative for Android, including the Attribution Reporting API (Android's equivalent of SKAdNetwork) and the deprecation of GAID over a multi-year window. The early Android operators we have worked with see GAID opt-out rates rising from the 18 to 28 percent range in 2022 to the 28 to 41 percent range in 2024, with the trajectory continuing. The Android attribution landscape in 2025 to 2026 is likely to look closer to the iOS landscape in 2022, with the deterministic identifier increasingly unavailable and the platform-provided aggregated APIs becoming the primary measurement layer.

Why Incrementality Has Become the Ground Truth

The combination of partial SKAdNetwork data, unreliable probabilistic attribution, the Android privacy trajectory, and the underlying impossibility of running a deterministic attribution layer over a non-deterministic user-identification regime has pushed mobile measurement programs toward incrementality testing as the operational ground truth. The shift is not new (incrementality has been the recommended ground truth for marketing measurement since at least the early 2010s), but it has accelerated sharply since 2021.

The incrementality framing reverses the attribution question. Attribution asks "which channel gets credit for this install?" Incrementality asks "if I had not run this campaign, would this install have happened anyway?" The two questions produce different answers: a campaign that targets users who would have installed organically gets credit under attribution but produces near-zero incremental installs under incrementality. A campaign that genuinely creates new installs gets credit under both. The decision-relevant quantity is the incremental count, not the attributed count.

The mechanics of incrementality testing on mobile have converged on geo-holdout designs as the dominant pattern. The test design: split a market (typically a country, sometimes a state or a DMA) into a treated group (campaign runs as normal) and a control group (campaign is held out, often by setting bid floors or geographic exclusions in the ad network). After a measurement window (typically 2 to 8 weeks), the install rate in the two groups is compared, and the difference is the incremental effect attributable to the campaign. The geo-holdout design works because the geography is randomly assignable (the geo split itself can be randomized across markets or randomized within a country), and the resulting comparison is causal in the sense that the only systematic difference between groups is the campaign exposure.

The diagram captures the workflow but understates the operational discipline required. The geo-matching step is critical: the treatment and control geos must be similar enough on the baseline install volume, trend, seasonality, and competitive context that the only systematic difference is the campaign. Mismatched geos produce confounded estimates that are worse than no estimate. The measurement window must be long enough to capture the campaign's full effect (including post-impression-decay effects, view-through effects, and any cross-channel ripple effects), but short enough that the macro environment does not shift underneath the test. The two to eight week window covers most app campaign dynamics; longer windows are required for high-consideration apps and shorter windows are insufficient for ramp dynamics.

The incrementality test produces an incremental install count and a per-incremental-install cost, both of which can be compared to the MMP-reported attributed install count and the MMP-reported cost-per-attributed-install. The ratio is the calibration: if the incremental count is 0.6 times the attributed count, attribution is overestimating effect by 40 percent. If the incremental count is 1.4 times the attributed count, attribution is underestimating effect or there is a measurement issue. The ratio varies by channel: in partner data, paid social channels typically run a 0.45 to 0.78 incremental-to-attributed ratio on app installs, search channels run a 0.58 to 0.92 ratio, video-network channels run a 0.32 to 0.68 ratio. The variation is the practical calibration coefficient that lets the attribution data be used as a directional signal for budget allocation.

The chart shows three values per channel because the variation matters as much as the median. The DSP retargeting bar (median ratio 0.34, P10 of 0.18) is a representative case: retargeting campaigns typically attribute installs to themselves that would have occurred anyway (the user was already interested, they would have re-engaged), and the incrementality testing reveals the overstatement directly. The Apple Search Ads bar (median ratio 0.84, P10 of 0.62) is at the other end: when a user searches the App Store for a specific app name or category, the ad is much closer to capturing incremental intent than displaying a banner to someone who was not searching.

The SAKA Problem and the Aggregated Attribution Layer

A specific failure mode worth naming is what the MMP industry has come to call the SAKA problem (the situation around Self-Attributing networks like Apple, Google, and Meta's properties versus the rest of the ecosystem). The Self-Attributing networks have direct visibility into both the ad serve and the eventual install, because they control both the ad-serving platform and (in Apple's case) the operating system, or because they have logged-in user identity that persists across the click-to-install path (Google, Meta). The non-SAN ad networks depend on the MMP-mediated SKAdNetwork postbacks and the probabilistic attribution layer.

The asymmetry creates a measurement gap. The SANs report attribution that includes installs the SKAdNetwork framework would not attribute to them (because the SAN has user-identity information SKAN does not), and the MMPs reporting the consolidated view either accept the SAN-reported number or apply heuristics to deconflict. The result is that the same install can appear to be attributed to multiple SANs, or to a SAN and a non-SAN, depending on which view is being read. The MMPs typically apply a deduplication logic, but the logic is heuristic and the result depends on the order of operations.

In partner data on operators running across SAN and non-SAN networks, the aggregated MMP-reported install count runs roughly 1.04 to 1.27 times the actual install count, with the over-count coming from the SAN over-attribution and the deduplication imperfections. The over-count is small enough at the aggregate level that it does not break the bigger picture, but it is large enough at the per-channel level that it can swing channel-level ROAS estimates by 10 to 25 percent.

The deduplication problem affects organic attribution most acutely, because organic is the residual category after every paid channel has been credited. If the paid channels collectively over-attribute by 15 percent, organic is correspondingly under-attributed by 15 percent of the same volume, and the operator's view of organic momentum is systematically wrong. The pattern is one of the underappreciated reasons app operators see organic growth that "should" be there but is not visible in the dashboard.

The remediation is partly methodological (the MMP and operator should agree on deduplication rules and document them) and partly architectural (the operator should run organic-incrementality tests periodically to validate the attribution layer's organic estimate). The organic-incrementality test is structurally similar to the channel-incrementality test: hold out paid activity in a test geo, measure the install rate, compare to the MMP-reported organic install rate in the same geo. The two numbers should be roughly similar; if the MMP's organic estimate is much lower, paid attribution is over-claiming credit that should be organic.

The 25-Conversion Floor and Long-Tail Strategy

The 25-conversion privacy threshold in SKAdNetwork has a specific operational implication that has reshaped how app operators design their paid campaigns: campaigns must be designed to clear the threshold or the value data is invisible. The threshold operates per attribution group (source app, campaign ID, conversion window), so a campaign that runs to a niche audience or a small geography or a narrow creative variation will not generate enough installs in the window to clear the threshold, and the conversion value comes back as null. The MMP can still see the install count, but cannot see the value, which means the ROAS cannot be computed for the campaign.

The strategic implication for campaign design is that small-budget tests and narrow targeting strategies have become unmeasurable on iOS for value-based metrics. The same campaign on Android (where the GAID is still available, for now) or in the IDFA-consenting iOS subsegment is measurable. The asymmetry produces a bifurcation in how operators run mobile measurement: aggregate budget into a small number of large campaigns that clear the threshold for the most important conversion values, accept that smaller campaigns will not produce value data, and use incrementality testing to validate the conversion-value estimates the aggregated data produces.

The bifurcation has reshaped media buying. The operators we have advised through this transition have consolidated their iOS paid media into fewer, larger campaigns; they have reduced the granularity of creative testing on iOS (because the conversion data is too sparse to differentiate creatives at the small-budget level); they have shifted creative testing to Android where the resolution is higher; and they have increased the frequency and scope of incrementality tests on iOS to validate the campaign-level value estimates.

How the Measurement Stack Should Be Structured in 2024

The 2024 measurement stack for a mobile app operator looks substantially different from the 2020 version. The 2020 version had MMP attribution as the single source of truth, with incrementality testing as an occasional validation. The 2024 version inverts the relationship: incrementality testing is the basis for budget decisions, MMP attribution is a directional signal that is calibrated against the incrementality results, and the SKAdNetwork data is one input to the MMP's attribution model rather than a standalone source.

The stack has four layers. Layer one is the platform-provided attribution APIs: SKAdNetwork on iOS, the Privacy Sandbox Attribution Reporting API on Android (as it matures), and the various SAN-direct integrations (Apple Search Ads Attribution API, Google Ads conversion API, Meta App Events API). The platform APIs provide the privacy-compliant aggregated attribution, with the limitations described above. Layer two is the MMP aggregation and normalization: AppsFlyer, Adjust, Branch, or equivalent, ingesting the platform APIs, applying probabilistic matching where the framework permits, deduplicating across networks, and producing the consolidated attribution view. Layer three is the operator's own analytics: in-app event tracking, customer data platform integration, post-install funnel analysis, and lifetime value calibration. Layer four is the incrementality testing program: regular geo-holdout tests on the major channels, periodic organic-incrementality validation, and channel-specific calibration coefficients applied to the MMP-reported numbers.

The four-layer structure has the property that no single layer is the ground truth, and budget decisions are made against the synthesis of all four. The MMP dashboard remains useful as a daily view of campaign activity, but the budget reallocation decisions are run against the incrementality calibration, not against the dashboard number directly. The discipline that distinguishes the mature measurement programs is the explicit acknowledgment of which layer answers which question, and the refusal to use the MMP number as the budget allocator without the calibration layer.

The operating cost of the four-layer stack is higher than the pre-IDFA single-source stack, both in headcount (the incrementality program requires statistical analysis and experiment-design skill) and in opportunity cost (the geo-holdouts withhold marketing spend in test regions, which is a real cost to the business). The cost is justified by the avoided cost of mis-allocated budget under the broken attribution regime: in partner data, the budget reallocation moves driven by incrementality-calibrated attribution typically produce 14 to 32 percent improvements in incremental-installs-per-dollar relative to the pure-MMP allocation. The improvement is enough to fund the program many times over for operators at any meaningful scale.

The two trajectories illustrate the practical economic case for the incrementality program. The pure-MMP allocation produces flat to slightly positive incremental-installs-per-dollar over the year because the MMP-attributed installs are not strongly correlated with actual incremental installs at the channel level. The calibrated allocation, where the MMP numbers are adjusted by the channel-specific incremental-to-attributed ratios and budget is reallocated accordingly, produces a steady improvement that compounds over the year as the calibration data accumulates and the allocation gets progressively closer to the incremental optimum.

The MMP dashboard answers "where did the install come from on paper." The incrementality test answers "would the install have happened without the campaign." For budget allocation, the second question is the only one that matters.

The Web-to-App Hybrid Path

A specific case worth treating separately is the web-to-app conversion path, where the user encounters the brand on web (typically via search or social), clicks a link, and either downloads the app from a deep link or eventually installs it through the App Store or Google Play. The web-to-app path was poorly served by the pre-IDFA attribution model (the cross-platform identification was always weak), and it is even more poorly served by the post-IDFA model.

The current state of web-to-app attribution: the web event (the ad click, the landing-page visit) is captured by the web analytics layer with a session identifier; the app install is captured by the MMP with whatever identifier is available (SKAdNetwork attribution, probabilistic match, or organic if neither resolves); the cross-platform stitch between the web session and the app install requires either a deterministic identifier (a logged-in user who completes the sign-in on both web and app, with the corresponding identity-resolution layer) or a probabilistic match (IP, geo, timestamp, device characteristics). The deterministic path is reliable but only covers logged-in users; the probabilistic path is unreliable at the individual-user level but produces aggregate estimates that are directionally useful.

For app operators where the web-to-app path is a meaningful share of installs (typical for consumer subscription apps, fintech apps, and any app with a strong web brand), the operational implication is that the web-to-app conversion rate has to be measured at the cohort level rather than the individual-user level. The standard approach is to estimate, for each web acquisition cohort (defined by source, campaign, geo, and date), the app-install rate over the following 7 to 28 days, and to attribute the installs to the web acquisition cohort based on the cohort-level statistical estimate rather than individual matching. The estimate is noisy but it is the best available signal, and the incrementality testing applies to the web-to-app path as much as to the direct-app-install path.

Key Takeaways

1. Last-click attribution on mobile apps stopped being operationally reliable in 2021 when iOS 14.5 broke the deterministic identifier chain. The IDFA opt-in rate settled at 18 to 32 percent for most apps, and the SKAdNetwork framework's structural limitations (postback loss, 25-conversion threshold, conversion-value noise) prevent it from filling the gap as a deterministic replacement.
2. The three SKAdNetwork limitations each have specific operational implications. Postback loss removes 11 to 36 percent of installs from network-level attribution. The 25-conversion threshold makes 18 to 42 percent of campaigns invisible at the value level. Conversion-value noise produces up to 12 percent error margins on individual postbacks that average out only at scale.
3. Probabilistic attribution restored some of the lost resolution but at the cost of regulatory risk and 64 to 84 percent match accuracy. Apple's position on probabilistic matching is contested and the enforcement could tighten without warning, which argues for treating probabilistic attribution as a transient mitigation rather than a stable measurement layer.
4. Geo-holdout incrementality testing has become the operational ground truth for mobile measurement. The design splits a market into treated and control geos, measures the install rate in both, and computes the incremental effect attributable to the campaign. The 2 to 8 week measurement window covers most app campaign dynamics.
5. The incremental-to-attributed install ratio is the practical calibration coefficient between MMP attribution and ground truth. In partner data, the ratio runs roughly 0.45 to 0.78 for paid social, 0.58 to 0.92 for paid search, 0.32 to 0.68 for video networks, and 0.18 to 0.54 for DSP retargeting. The variation is the practical input to incrementality-calibrated budget allocation.
6. The SAKA problem (the asymmetry between self-attributing and non-self-attributing networks) produces 4 to 27 percent over-attribution at the aggregate level and larger swings at the per-channel level. The deduplication problem disproportionately affects organic attribution, which gets systematically understated.
7. The 25-conversion threshold has reshaped campaign design. Operators have consolidated iOS campaigns into fewer, larger structures that clear the threshold; small-budget tests and granular creative experiments have shifted to Android where the resolution is still higher; the incrementality testing layer has expanded to validate the consolidated structure.
8. The 2024 mobile measurement stack has four layers (platform APIs, MMP aggregation, operator analytics, incrementality testing) and the budget decisions are made against the synthesis of all four. The incrementality-calibrated MMP attribution typically produces 14 to 32 percent improvements in incremental-installs-per-dollar relative to the pure-MMP allocation, which is large enough to fund the program at any meaningful scale.

Citations and Further Reading

• Apple Developer Documentation, SKAdNetwork API Reference (versions 2, 3, and 4), the canonical reference for the framework's mechanics, conversion windows, and privacy thresholds.
• Apple's App Tracking Transparency framework documentation and the published App Store Review Guidelines, including the prohibitions on fingerprinting and the requirements for the ATT prompt.
• AppsFlyer State of App Marketing reports (annual), including the post-iOS-14.5 measurement adaptation guidance and the published opt-in rate trends by vertical.
• Adjust's published guides on SKAdNetwork campaign structure, conversion value design, and the practical operating mechanics of the framework.
• Branch's documentation on web-to-app attribution, deep linking, and the cross-platform identity-resolution problem.
• Singular and Kochava's published research on probabilistic attribution accuracy and the validation methodologies against the IDFA-opt-in sample.
• Google Privacy Sandbox for Android, the public documentation on the Attribution Reporting API and the GAID deprecation roadmap.
• Meta's published documentation on the App Events API, the SAN-side attribution methodology, and the published guidance on the SKAdNetwork integration for Meta-managed campaigns.
• The Mobile Measurement Partner industry conferences (MAU, App Promotion Summit, others) for the published practitioner discussions of incrementality testing methodology and the calibration patterns across verticals.
• The academic literature on causal inference for media measurement, including Kay Brodersen et al. (Google, 2015) on Bayesian Structural Time Series for incrementality testing and the broader Imbens and Rubin (2015) treatment of causal inference design.
• The published case studies on geo-holdout testing from Meta's Marketing Mix Modeling team, Google's experiment-design publications, and the industry adaptations by major DTC and gaming operators.
• Eric Seufert's published commentary on mobile measurement, the SKAN transitions, and the operating implications for app operators in the post-IDFA era.