A mobile app marketing audit is not a scorecard. A scorecard assigns points to a list of practices and produces a grade that nobody can act on. An audit is a decision system: it establishes what decision is pending, what evidence is authoritative for that decision, what the evidence currently supports, and what must change before money moves.
That distinction matters most at the moment teams usually ask for an audit. Spend is working well enough to continue and badly enough to worry about. Cost per install rose. Platform-reported return fell. Retention looks thinner than last quarter. The team has six dashboards and no agreement about which one is right. The temptation is to optimize media, because media is the surface with the most buttons. Frequently the binding constraint sits somewhere else entirely — the store listing, the event schema, the onboarding path, the monetization moment, or the definition of the number everyone is arguing about.
This guide sets out an evidence-first audit method for paid app acquisition: how to write the audit contract before touching data, how to walk a six-layer evidence ladder, how to label the confidence of every finding, how to separate a symptom from its owner, how to convert findings into Keep, Fix, Constrain, or Stop decisions, and how to prioritize work when several findings compete. It also states plainly where audits cannot reach.
What a mobile app marketing audit actually is
An audit answers one question: given the evidence we can trust today, what is the responsible next decision about this growth system?
Three properties separate it from a checklist review.
It is scoped to a decision. "Should we increase paid budget by 40% next quarter?" is a decision. "Is our marketing good?" is not. The decision determines which evidence is relevant and which is decoration.
It ranks evidence, not opinions. Every finding carries a source and a confidence state. A platform dashboard, a store analytics export, an attribution provider report, and an internal revenue ledger can all describe the same week and disagree. An audit names which source has authority for which claim before it reads any of them.
It ends in decision rules, not adjectives. "Creative is weak" is an adjective. "Concept family B has not produced a mature payer cohort at or below the CAC ceiling across 60 days and two territories; constrain to the one placement where it reconciles, or stop it" is a decision rule.
The rest of this article assumes the paid system is already running or about to scale. If you are still designing channel, event, and creative structure, the mobile app UA strategy framework covers that ground, and budget sizing belongs in the mobile app marketing budget guide.
Step one: write the audit contract
Most audits fail before analysis begins, because the participants never agreed on terms. Write the contract first, in one page, and get explicit sign-off from whoever owns the budget.
The contract fixes ten things.
1. The decision. One sentence. Scale, hold, restructure, repair instrumentation, or stop a specific investment.
2. The business outcome that decides it. Not a proxy. Contribution after variable cost, payback period, retained subscribers, or qualified activations — whatever your finance partner will actually defend.
3. The owner. The person who can act on the conclusion. An audit delivered to nobody in particular produces nothing.
4. The date range. Fixed start and end, stated in one timezone, with the reason for the boundaries.
5. Cohort maturity. The minimum observation window for the outcome in question. If payback is judged at day 30, cohorts acquired 11 days ago cannot be judged at all. They are not bad; they are not mature.
6. The source of truth per claim. Store conversion comes from store analytics. Spend comes from the ad platform billing view. Revenue comes from the financial ledger or the store's proceeds reporting. Attribution comes from the attribution provider or the platform, with the definition written down.
7. The attribution definition. Click window, view-through treatment, modeled conversion treatment, re-attribution and redownload rules, and how each channel differs. Apple Ads, Google Ads, and paid social define acquisition differently by design. The contract records the differences instead of pretending they do not exist.
8. Excluded traffic. Internal testing, QA devices, employee installs, incentivized inventory, and known bot patterns.
9. Known events. Launches, price changes, store listing updates, SDK upgrades, OS releases, seasonal peaks, PR spikes, and outages inside the window. An unannounced pricing test explains more variance than most creative reviews.
10. Stop criteria. The conditions under which the audit halts and reports an instrumentation problem instead of a media conclusion. This clause is the one most often omitted and the one that saves the most money.
The evidence ladder: six layers, bottom to top
Diagnose in order. A finding at a higher layer is unreliable while a lower layer is broken, because every upper layer inherits the lower layer's definitions.
Layer 1 — Store discovery and conversion
Questions. How many people reach the product page from paid traffic, and how many install? Does store conversion differ by traffic source, territory, device, and listing variant? Did a listing change coincide with the shift in cost per install?
Authoritative evidence. Store analytics. Apple documents how App Store Connect analytics reports acquisition and which dimensions and filters exist, including the distinction between different acquisition sources (App Store Connect analytics overview; acquisition definitions; filters and dimensions). Google documents store-listing performance reporting and its acquisition and retention reports for Play (store listing performance; acquisition reports).
Common false positive. Concluding that media quality collapsed when paid traffic simply started landing on a page that converts less well for that audience. Cost per install is a compound of auction price and page conversion; the auction gets blamed because it is visible in the ad account.
Exit criterion. Store conversion for the relevant source, territory, and device is measured over a stable window, and any listing change inside the window is documented. Page-level testing detail belongs in the App Store screenshot and page-conversion guide.
Layer 2 — Paid media delivery and creative promise
Questions. Where did delivery actually go — placements, territories, audiences, formats? What promise does each creative family make, and does the store page keep it? Is the campaign optimizing toward the event the business cares about?
Authoritative evidence. Platform delivery and asset reporting, alongside the platform's own documentation on goals, assets, and stabilization. Google publishes App campaign setup by goal, best practices, and tips that address conversion delay and asset requirements (setup and goals; best practices; tips). Apple Ads publishes its reporting options and metric definitions and its Insights reports (reporting definitions; Insights reports).
Common false positive. Reading a "top performing" asset label as causal. Assembly-based systems report components of impressions that were served together; a component appearing in winning combinations is not proof it caused the outcome. Goal-selection questions are handled separately in the Google App campaign goals guide, and keyword-level structure in the Apple Ads keyword strategy guide.
Exit criterion. Delivery composition is known, the optimization event is named, and creative families are grouped by the promise they make rather than by file name.
Layer 3 — Attribution and event integrity
Questions. Does the event fire, once, with the right parameters, on both platforms, in every build currently in the wild? Do platform-attributed conversions reconcile approximately with the attribution provider and with the revenue ledger? Are modeled and view-through conversions separated from directly observed ones?
Authoritative evidence. Conversion tracking configuration and the analytics event schema. Google documents mobile app conversion tracking and a recommended event taxonomy for apps (app conversion tracking; recommended app events).
Common false positive. Treating a zero as a fact. A revenue column of zero can mean no revenue, or no purchase event implemented for that build, or a parameter type mismatch, or a privacy threshold suppressing the row. These are different problems with different owners. Deeper architecture and decision-rights questions belong in the attribution architecture guide.
Exit criterion. Each business-critical event has a named owner, a documented trigger, a verified payload, and a known coverage gap list by app version and platform.
Layer 4 — Onboarding and qualified activation
Questions. What proportion of installs from each acquisition cohort reach the first genuinely useful moment, not merely the end of the tutorial? Where do cohorts diverge, and does the divergence match the creative promise made upstream?
Authoritative evidence. In-product analytics with a defined activation event, segmented by acquisition source and campaign.
Common false positive. High onboarding completion read as success. Completion measures whether people finished your screens. It does not measure whether they received value. A cohort can complete onboarding at a high rate and retain poorly, which usually indicates the flow is short rather than effective. The full diagnosis lives in the onboarding optimization guide.
Exit criterion. A single activation definition exists, is applied identically across channels, and is old enough to be observed for the cohorts under review.
Layer 5 — Retention and monetization by mature cohort
Questions. For cohorts old enough to judge, what does retention look like by acquisition source? What is realized revenue per cohort, split by subscription, in-app purchase, and advertising? How much of the reported value is observed and how much is forecast?
Authoritative evidence. Store retention reporting, subscription state reporting, and the internal financial ledger — with Apple's and Google's differing definitions noted rather than blended (Play acquisition and retention; App Store Connect analytics).
Common false positive. Comparing an immature cohort to a mature one and calling the difference a decline. Value calculation method is covered in the mobile app LTV framework.
Exit criterion. At least one cohort per material channel has reached the maturity threshold named in the contract, with observed value separated from modeled value.
Layer 6 — Incrementality and economic reconciliation
Questions. Would some of this outcome have happened anyway? Does total spend reconcile with total contribution at the business level, not only at the campaign level?
Authoritative evidence. Designed holdout or geographic tests, plus finance-side reconciliation. Attributed conversions cannot establish incrementality no matter how many platforms agree, because they all measure exposure-adjacent outcomes rather than counterfactuals. Test design is covered in the incrementality testing guide.
Common false positive. Reading channel-reported return as a lift estimate.
Exit criterion. Either a valid incrementality read exists, or the audit states clearly that incrementality is UNAVAILABLE and confines its conclusions accordingly.
Label every finding with an evidence state
Findings without confidence labels get flattened into a single tone of authority in the final read-out. Use five states and use them literally.
State Meaning Allowed use --- --- --- VERIFIED Reconciles across the authoritative source and at least one independent source Can support a budget decision DIRECTIONAL One credible source, not reconciled, or a small sample Can support a test, not a budget decision CONTRADICTORY Two credible sources disagree materially Blocks the decision until resolved UNAVAILABLE The data does not exist, is suppressed, or was never implemented Must be reported as a gap, never as a zero NOT MATURE The mechanism works but the observation window is too short Reported with the date the read becomes possible
The distinction between UNAVAILABLE and a genuine zero is the single highest-value discipline in this method. Privacy thresholds suppress small rows. Opt-in and consent conditions restrict what some sources can report at all. Apple's app analytics documentation describes the privacy conditions and opt-in behavior that shape what appears in those reports (app analytics overview). A suppressed row and an empty result look identical in a spreadsheet and mean opposite things.
Platform measurement also changes. Google documented a measurement change to store listing performance reporting effective July 2026, which affects how listing-level performance is counted and therefore how earlier and later periods compare (store listing performance). Any year-over-year store comparison that straddles a definitional change should be labeled CONTRADICTORY until it is normalized.
Separate the symptom from the owner
Every headline metric has several plausible owners. Assigning the wrong owner is how teams spend a quarter fixing the wrong layer.
Symptom Possible owners First evidence to pull --- --- --- Cost per install rose Store conversion fell; creative fatigue; audience or placement shift; auction pressure; tracking gap inflating cost per recorded install Store page conversion by source and territory, then delivery composition Platform-reported return fell Delayed revenue not yet observed; attribution window change; event payload broken; genuine monetization decline Conversion lag curve and event payload check before any media change Installs up, revenue flat Promise/product mismatch; low-intent placements; activation failure; monetization gating Activation and payer rate by acquisition cohort Onboarding completion strong, retention weak Onboarding too shallow to deliver value; wrong activation definition; audience mismatch upstream Cohort retention by creative family, not by campaign One channel looks dramatically superior Different attribution definitions across channels; last-touch credit for demand created elsewhere Definition comparison, then a holdout
None of these mappings is automatic. They are hypotheses that the evidence ladder confirms or rejects.
The issue decision model: Keep, Fix, Constrain, Stop
Every confirmed finding resolves into exactly one of four outcomes. These are outcomes of this audit under this contract — not universal recommendations.
Keep. The mechanism's role is clear and the economics reconcile against the business outcome at the stated maturity. Keep does not mean untouched; it means no decision is pending.
Fix. The mechanism is appropriate but the evidence or implementation is faulty. Broken event payloads, misdefined activation, an untested listing, or a conversion window that contradicts the buying cycle. Fix items usually block other findings and therefore go first.
Constrain. The mechanism produces acceptable outcomes only within a narrow boundary — a cohort, a territory, a placement, a creative family, or a budget ceiling. Constrain is often the correct answer for a campaign that looks bad in aggregate and reconciles cleanly in one segment.
Stop. After the evidence is corrected, the mechanism still cannot meet the decision in the contract. Stop is only defensible after Fix has been attempted on any finding that was CONTRADICTORY or UNAVAILABLE, because stopping on broken evidence destroys working spend.
Prioritizing when several findings compete
A real audit produces more findings than any team can act on in one cycle. Rank them on six dimensions. Do not apply fixed universal weights — the weights depend on the decision in the contract and on how much capital is at risk.
Evidence confidence. VERIFIED findings outrank DIRECTIONAL ones for budget decisions.
Spend exposed. How much money flows through the mechanism per week while the finding remains unresolved.
User harm. Whether the issue degrades the experience of real users, not only reporting accuracy.
Reversibility. Cheap-to-undo changes can proceed under lower confidence than irreversible ones.
Time to learn. How long before the fix produces an observable, mature signal.
Dependency order. Whether other findings are unreadable until this one is resolved. Instrumentation fixes almost always sort first for this reason alone.
A practical rule: resolve every dependency-blocking Fix before acting on any Stop, and never release additional budget on a DIRECTIONAL finding.
A responsible audit workflow
Freeze definitions. Publish the contract. Nobody redefines a metric mid-audit.
Collect raw exports. Platform, store, attribution, product analytics, and finance — as exports, not screenshots, with generation timestamps.
Reconcile denominators and windows. Align date boundaries, timezones, attribution windows, and currency before comparing anything.
Segment by cohort. Acquisition cohort, not reporting week. Weekly reporting mixes mature and immature users into one misleading average.
Map contradictions. List every place two authoritative sources disagree, with the size of the gap. Resolve or label.
Form hypotheses. One sentence each, with the layer that owns it and the evidence state.
Choose the smallest test. The cheapest change that would distinguish between the two leading hypotheses.
Define pass, fail, and stop rules in advance. Including the date the read becomes valid.
Release budget only after the downstream signal matures. Not after the platform metric improves.
A labeled fictional example
All numbers below are invented for illustration. They are not benchmarks, industry averages, or Sharply Labs client results.
A fictional app spends $100,000 in a month across two channels.
Channel A Channel B --- --- --- Spend $60,000 $40,000 Installs 30,000 10,000 Cost per install $2.00 $4.00 Qualified activation rate 8% 26% Qualified activations 2,400 2,600 Cost per qualified activation $25.00 $15.38
Channel A's install is half the price. Channel B produces qualified activations at roughly 62% of Channel A's cost. A media review that stops at cost per install would shift budget in exactly the wrong direction.
Extend it one layer. Suppose 6% of Channel A's activations and 11% of Channel B's become mature payers by day 30:
Channel A: 144 mature payers, $416.67 each.
Channel B: 286 mature payers, $139.86 each.
Now add the failure that makes this example realistic. Suppose the purchase event on the Android build shipped with a malformed value parameter, and Channel A's traffic is 80% Android while Channel B's is 30% Android. The payer counts above are not comparable, and neither is any revenue figure derived from them. The correct audit output is not "shift budget to B." It is: event integrity is CONTRADICTORY; the payer comparison is blocked; Fix the payload, re-observe one mature cohort, and hold the budget decision until then. The activation comparison, which does not depend on the broken parameter, remains usable as DIRECTIONAL evidence.
That is the whole point of the ladder. The cheapest conclusion available was wrong twice, in opposite directions, before the evidence was repaired.
When not to run an audit
When no decision is pending. Audits consume senior attention. Without a decision, the output is a document.
When no cohort is mature. If the app launched three weeks ago and the outcome is judged at day 30, you can audit instrumentation and store conversion, and nothing else. Say so rather than producing early reads dressed as conclusions.
When the real requirement is instrumentation repair. If half the critical events are unverified, the honest deliverable is a repair plan, not a media optimization plan. Optimizing against an unreliable signal teaches the delivery system the wrong lesson and the damage compounds while the report reads normally.
When the commercial question is actually an operating-model question. Whether to hire, outsource, or restructure is a different question from this one, and an audit should say so rather than answering it by implication.
Buyer questions, answered directly
What is included in a mobile app marketing audit? A written contract, a layer-by-layer evidence review across store, media, measurement, activation, retention, and economics, a labeled findings list with evidence states, symptom-to-owner mapping, Keep/Fix/Constrain/Stop decisions, a prioritized sequence, and the specific next test with its pass, fail, and stop rules.
How is a paid UA audit different from an ASO audit? An ASO audit examines organic discovery and store-page performance in its own right. A paid UA audit treats the store page as one layer inside the paid path — specifically, as the conversion step between the ad click and the install — and extends through attribution, activation, retention, and economics. They overlap at the store listing and diverge everywhere else.
Which data should an agency request? Ad platform exports with spend and delivery detail, store analytics for acquisition and conversion, attribution provider exports with the window configuration, product analytics with the activation and monetization event schema, and a finance-side revenue view. Aggregated and cohort-level exports are sufficient; no personal data should change hands.
How do you audit across Apple Ads, Google Ads, and paid social without forcing identical attribution definitions? You do not force them. Record each platform's definition, compare each channel against itself over time, reconcile all channels against one shared downstream outcome such as qualified activations or mature payers from the product analytics layer, and use holdouts where a cross-channel causal claim is required. Apple Ads' reporting definitions and Google's app conversion tracking documentation are the reference points for what each platform is actually counting (Apple Ads reporting definitions; Google app conversion tracking).
How often should an app marketing audit be run? Full audits are triggered by decisions, not by calendars: before a material budget increase, after a significant product or pricing change, when sources start disagreeing, or when the outcome moves without an explanation. A lightweight instrumentation check on a regular cadence is different and worth keeping.
What should the deliverable contain? The contract, the findings table with evidence states, contradictions with their size, the decision for each finding, the prioritized sequence with owners, the next test with its rules and read date, and an explicit list of what could not be verified and why.
Can an audit guarantee lower CPI, CPA, or higher ROAS? No. An audit improves decision quality and reveals testable constraints. It cannot guarantee a cost or return outcome, because outcomes depend on auction conditions, product, competition, pricing, and seasonality that no diagnostic controls. Any provider promising a specific performance number from a diagnostic is describing a sales position rather than a method — a useful thing to probe when evaluating a mobile app marketing agency.
Limitations you should state out loud
Privacy thresholds suppress small segments; absence of a row is not absence of activity.
Opt-in bias means some reported populations are not representative of all users.
Modeled and view-through conversions are estimates and should never be blended silently with observed conversions.
Redownload and re-attribution definitions differ by platform; a "new user" is not the same entity everywhere.
Conversion lag makes recent windows systematically understate outcomes.
Cohort immaturity produces pessimistic reads that look like decline.
Cross-platform definitional differences are permanent and must be managed, not resolved.
Attributed conversions cannot establish incrementality. Only designed tests can, and even then only within their scope.
The minimum viable evidence pack
Assemble this before any diagnostic conversation. It contains no personal data.
The pending decision, in one sentence, and the date it must be made.
The business outcome that decides it, with its current definition.
A fixed date range and timezone.
Spend and delivery exports by channel, campaign, territory, and platform.
Store analytics: product page views, conversion, and acquisition source breakdowns.
Attribution configuration: windows, view-through treatment, re-attribution rules, per channel.
The event schema for install, activation, and monetization events, with app versions and known gaps.
Cohort retention and realized revenue for at least one mature cohort per material channel.
A dated list of launches, price changes, listing updates, and outages in the window.
A written note of everything you already know is unreliable.
Teams that arrive with items 1, 2, 9, and 10 are usually two weeks ahead of teams that arrive with only dashboards.
---
If your app is already spending, or preparing to scale, and store, media, attribution, activation, retention, and monetization evidence will not reconcile, a short conversation is usually enough to find out where the system is actually blocked. Our mobile app growth practice and performance marketing services start from the same place this article does.
Book the free 15-minute conversation. We will work through your audit contract, the definitions currently in conflict, the evidence gaps that block a decision, and the next smallest reliable test. You will leave with a scoped diagnostic discussion and the specific evidence list required to reach a defensible conclusion. We will not promise a lower cost per install, a higher return, or better retention — those depend on your product, market, and auction, not on a diagnostic.