The mobile UA market in 2026 is not the one your bidding models were built for. Global app downloads grew just 0.8% in 2025, according to Sensor Tower's State of Mobile 2026, which is the structural confirmation that the era of volume-led growth is over. We covered the broader trend lines from that report in Mobile Game User Acquisition Trends 2026, but those trends raise a measurement problem that the report itself does not solve.
When the cost of a bad acquisition decision compounds faster than it used to, when channel concentration has shifted decisively toward Meta, and when the gap between ad spend and revenue in entire categories has widened past 14 percentage points, the question of whether your campaigns are producing real lift or just claiming credit for installs you would have got anyway becomes the question that determines whether your UA budget is profitable.
Incrementality testing is how you answer that question. Most teams running mobile UA at scale today are not doing it, and the cost of not doing so is now higher than ever
Incrementality testing answers a single question. If you turned off a campaign, channel, or audience tomorrow, how many of the conversions it currently reports would you still get? The portion you would lose is incremental. The portion you would not lose was always going to happen, and you paid for it.
This is a different question from attribution. Attribution decides who gets credit for a conversion that already occurred. Incrementality determines whether the conversion occurred because of the ad.
The two answers diverge most in retargeting, branded search, lookalike audiences, and any optimization signal that rewards the platform for finding users who were already going to convert. Self-reported ROAS from any walled garden tells you what the platform claims credit for. Incrementality tells you what the platform actually caused.
Three findings from the report directly raise the stakes on whether your UA decisions are calibrated to causality or to credit-claiming.
The first is the spend-to-revenue mismatch by category. In the US in 2025, Lifestyle & Puzzle attracted 56.2% of total gaming ad spend while generating only 41.3% of IAP revenue. Casino sat at the inverse, 10.4% of spend capturing 21.6% of revenue. Action & Strategy ran near balance at 30.7% spend versus 34.8% revenue.
A category running 14 points of ad spend above its revenue weight is a category where auction competition has driven CPMs above what the revenue pool can support. In that environment, the difference between paying for incremental users and paying for users who would have installed anyway is no longer a measurement nicety. It is the variable that decides whether the campaign is profitable.
The second is the regional revenue divergence. Asian publishers grew IAP revenue by $2.58 billion in 2025. North American publishers declined by $1.78 billion. The deeply monetized mid-core titles driving Eastern growth generate higher LTV per user, which means Eastern UA teams can profitably bid higher than Western competitors in the same auction. If your product produces $2 of LTV per install and a competitor's app produces $5, they can pay $3.50 and win. You cannot follow them there without losing money. The recoverable efficiency for Western studios in this environment is not in outbidding. It is in not paying for installs that were never incremental in the first place.
The third is the casual retention collapse. Casual D1 retention among top-25 revenue titles now sits at approximately 30%, below mid-core at 44 to 45% and below hypercasual at 38 to 40%. D7 has fallen to roughly 14 to 15% for casual, beneath even hybrid-casual. D30 sits at 7 to 8%. The compounding effect across three retention windows is the issue: an LTV model calibrated on 2022 or 2023 retention curves is overestimating cohort value at every milestone, which means the CPI you are willing to pay is calibrated to revenue your cohort is no longer producing. Incrementality testing is one of the few tools that can tell you whether the gap between your model and your reality is being papered over by platform credit-claiming or whether you are actually buying lift.
Three test designs do most of the practical work. Each has trade-offs, and the right choice depends on what you are testing, where you’re spending, and how your data infrastructure is set up.
Split a market into matched regions, run ads in one set, hold spend out of the other, and compare aggregate installs and revenue. This is the structural backbone of channel-level lift studies and the natural fit for post-ATT iOS because it does not require user-level identifiers.
Geo tests work for evaluating whether a whole channel, or a major spend reallocation, produces real lift. The design constraints are real. You need enough geographic separation to create a defensible control, typically 10% to 20% of geography held out, and enough volume in those regions for the signal to clear natural market noise. Test windows usually run two weeks for fast-converting categories and four to six weeks for considered-purchase verticals.
A randomly selected portion of the eligible audience is excluded from seeing your ad. The conversion rate of the test group minus the conversion rate of the control group, divided by the test group, gives you incremental lift. This is the design behind the managed lift products at AppsFlyer, Singular, Adjust, and the platform-native conversion lift tools at Meta and Google.
User-level holdouts deliver tighter statistical confidence at smaller sample sizes than geo tests. The trade-off on iOS is that holdouts depend either on ATT-opted-in users or on platform-managed audiences, both of which carry their own coverage and bias issues. For Android, and for retargeting on iOS, this is generally the cleanest design available.
Pause a campaign for a defined window and measure the difference in conversion volume against the matched periods before and after. The cheapest test to run, the easiest to misinterpret. Seasonality, competitor activity, app store featuring, and creative fatigue can all swamp the signal in either direction. Time-based holdouts are useful as sanity checks. They are not the basis for a budget decision.
A workable internal process runs in roughly four stages.
Pre-test design starts with a hypothesis specific enough to fail. Not "is Meta working" but "does our Meta retargeting spend on lapsed payers produce more than $X in incremental D7 revenue compared to control." From there, run a power analysis. Most managed lift tools and the open-source GeoLift R package expose this calculation directly. If the test is underpowered before it starts, kill it and either consolidate spend until volume is sufficient or pick a different test target.
Execution requires holding the conditions constant. No creative refreshes mid-test, no budget shifts, no targeting changes. Document every external factor that could move the signal. App store features, PR moments, competitor launches, and seasonality all need to be in the test log so anomalies can be explained later.
Analysis is where most teams cut corners. Calculate incremental conversions and incremental revenue, then convert those to incremental cost per acquisition (iCPA) and incremental ROAS (iROAS). Statistical significance matters. A 15% lift estimate with a 30% confidence interval is not a result. Report the lift, the confidence interval, and the cost of the holdout itself.
Decision and rollout are the parts most UA teams skip. The test outcome should produce a specific budget action: scale, hold, cut, or restructure. Document the decision and the reasoning. The next time someone questions a channel allocation, the test log is the answer.
Three categories show up consistently when studios start running their first round of lift tests.
Branded search and brand-bid retargeting almost always over-credit themselves. Users searching your game by name were going to install regardless of whether you bid on the keyword. Lift tests in this category routinely return iROAS that is a fraction of platform-reported ROAS, sometimes near zero.
Lookalike audiences built on small or noisy seed sets convert at rates that look strong in attribution dashboards but produce minimal lift in controlled tests. The platform finds users who would have installed organically and claims them.
Over-targeted high-intent audiences in mature markets exhibit the same pattern. The optimization signal rewards the platform for finding users who were already going to convert, which means the platform increasingly does that.
The connection back to the State of Mobile 2026 data is direct. If you are operating in Lifestyle & Puzzle in the US, where 56.2% of ad spend is chasing 41.3% of revenue, the over-credit on retargeting and branded campaigns is what determines whether your unit economics work. The campaigns that look profitable on platform-reported numbers may be the ones absorbing the auction premium without producing the lift to justify it.
Incrementality is not a replacement for attribution or for SKAdNetwork. The 2026 measurement stack is layered. SKAN postbacks for deterministic-anonymous reporting on opted-out iOS. MMP attribution and probabilistic modeling for opted-in cohorts. Media Mix Modeling for the long-horizon view that incorporates seasonality, brand, and offline factors. Incrementality testing is the cross-check that tells you which of the above is closest to true at specific points in the spend curve.
The teams scaling UA profitably in 2026 are not picking one of these methods. They are running all of them and reconciling them when they disagree. Incrementality is what makes the reconciliation possible.
The right first test is small, focused on a channel where you already suspect over-credit, and designed to produce a clear yes-or-no answer about a specific spend slice. Retargeting on a high-spend channel is usually the right place to start because the suspected over-credit is large and the cost of holding out is low. Branded search is the next most useful starting point.
Avoid testing your largest top-of-funnel campaign first. The cost of error is high, the variables are many, and team confidence in the methodology is still building. Start where the answer is most likely to change a decision and least likely to disrupt revenue if the test goes sideways.
Budget at least one full retention cycle for the test window. For most mobile games, that is four weeks minimum to capture meaningful post-install behavior. Less than that, and you are measuring installs, not value. Given how steeply the casual retention curves have declined, four weeks is also the window where the gap between modeled and actual cohort value becomes visible, which makes it the right horizon for a test designed to validate LTV assumptions.
The shift from attribution-led UA to incrementality-validated UA is happening unevenly. Large publishers and well-resourced studios have been running structured lift tests for years. Mid-size studios and most independent developers are still optimizing primarily on platform-reported ROAS and SKAN postbacks, which means they are funding the gap between reported lift and real lift out of their own margins.
For studios competing in auctions where Eastern publishers can profitably bid higher, that gap is not a soft cost. It is the difference between a profitable quarter and a missed one. Incrementality testing is one of the few tools that can change what the scoreboard says about your team next year.
If your team is navigating the 2026 UA market and could use a second set of eyes on strategy, measurement, or execution, GameBiz Consulting works with mobile studios across the full UA stack. Get in touch before the next quarterly plan locks in.
Source: Sensor Tower, State of Mobile 2026. Trend analysis adapted from GameBiz Consulting, Mobile Game User Acquisition Trends 2026: What the Data Says.
