Why a web plan does not carry over to apps
On the web a tag manager lets an analyst fix tracking in an afternoon. In an app, the event specification is part of the product build: it goes through code review, a release train and store review, and then reaches users only as fast as they update. A plan written after launch therefore inherits whatever the first version happened to record, sometimes for months.
The second difference is the number of counters. The store console, each ad network, the mobile measurement partner and the in-app analytics tool all report installs, each with its own definition, time zone and attribution rules. The third is privacy. Where user-level matching is restricted, the ad networks learn about post-install behaviour through a compressed signal designed by the advertiser, and a plan that does not design that signal leaves the networks optimising blind. The app marketing hub frames the install-to-retention loop the plan has to serve; the plan makes each stage of that loop countable.
Decisions that change for app installs
| Plan component | In the generic plan | In an app-install plan | Decision it forces |
|---|---|---|---|
| Tracking specification | Tags that can change any day | Events compiled into SDK releases | What must be specified before the first campaign version ships |
| Source of truth | One analytics property | A measurement partner, an analytics tool, networks and stores | Which tool owns installs, which owns behaviour, which owns spend |
| KPI for paid media | A conversion event with full attribution | Installs plus a coarse post-install signal on restricted devices | How conversion values encode early value |
| Quality checks | Tag QA on pages | Link routing across installed and new users | Which deep-link paths are tested every release |
| Reconciliation | Rarely needed | Store, partner and analytics counts compared | What gap is tolerated and who investigates beyond it |
| Owners | Analyst and developer | Product engineering, growth, analytics and the partner account | Who approves an event change in the release |
Specify events before the release that needs them
Because tracking ships with the product, the plan has to be ahead of the build. List the install definition, the activation event, the first purchase and the parameters each carries, and put them in the sprint before the campaign rather than in the campaign brief. The cost is upfront effort for events that may turn out unimportant; the alternative is a campaign that runs for weeks on a version that cannot report its own KPIA metric tied directly to an objective.. Choose to specify a slightly larger set than seems necessary at first launch, and prefer events that can also be sent from the server, since those can be corrected without an app update. The deep links and app analytics guide covers naming an event taxonomy for apps; the plan decides which of those events prove which goal.
Decide what each tool owns
The measurement partner is good at deciding which campaign earned an install and at forwarding 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. to networks. The app analytics tool, for example GA4Google's event-based analytics platform. fed by a Firebase SDK, is better at behaviour inside the app. The networks own spend and delivery. Overlap is where arguments start: if two tools both report "installs by campaign", someone will compare them. Write down which tool answers which question, and switch off or ignore the duplicate. Choose the partner as the source for paid AttributionAssigning conversion credit to touchpoints., the analytics tool for product behaviour and cohorts, and the finance system for cost. The guide to app attribution, MMPs and SKAdNetwork explains what each layer can and cannot see.
Design conversion values on purpose
Where attribution is privacy-limited, the advertiser chooses what a small number of coarse values mean and how long the app waits before reporting. Report early and the value says little; wait longer and the signal arrives too late to steer spend. Encode the behaviour that best predicts the business outcome within the window available, such as completing a first order, rather than a vanity action such as viewing a screen. Choose a schema that a person can explain in one line, document it beside the event spec, and date every change, because each change breaks comparison with earlier postbacks.
Put deep-link tests in the QA checklist
Install campaigns often promise a specific screen: a discount, a category, a referral reward. If the link drops a new user on the home screen, conversion falls and the plan's activation metric reports a product problem that is really a routing bug. Add each campaign's deep link to the release QA list, tested for users who already have the app and for new installs through deferred linking. The trade-off is QA time on every release against silent leaks that only appear in cohort data weeks later; choose the test. The deep link testing tutorial sets out that test step by step, from the route table to the deferred path on a fresh device.
Treat store figures as a reconciliation, not a rival
Store consoles count downloads by their own rules, including re-installs and devices that never open the app. They will not match the partner or the analytics tool, and they should not be expected to. Write the expected gap down as a known difference, set a tolerance, and name who investigates when it is breached. The tutorial on reconciling app install numbers builds that monthly comparison, sets the tolerance from the team's own history and investigates when a gap moves. The app install campaigns guide covers optimising campaigns once the counts are trusted.
Decision points in a worked app-install plan
A measurement plan for an invented grocery app, Basketful
Basketful delivers groceries in three cities and is launching a paid install campaign with a budget of £60,000 over eight weeks across two networks. Its objective is "build a base of weekly ordering customers". The plan, condensed:
| Goal | KPI | Owner tool | Event or source | Target | Release needed |
|---|---|---|---|---|---|
| People install and open | First opens from paid media | Measurement partner | first_open | 24,000 in eight weeks | Already live |
| New users place an order | First order within 7 days of open | Analytics tool | purchase with order_number = 1, also sent server-side | 30% of opens | Version 4.12, two weeks before launch |
| Paid networks see early value | Conversion value from first 48 hours | Measurement partner | Values: opened, browsed a category, basket over £20, ordered | Coarse value "ordered" on 18% of postbacks | Schema signed off at launch |
| Promised screen reached | Deep-link success rate | QA and analytics | deep_link_open with target_screen | Over 95% on both platforms | Version 4.12 |
| Counts reconcile | Partner opens against store first-time downloads | Analytics lead | Weekly comparison | Gap within 15% | n/a |
Cost per first order is the headline: £62 at target, from £58 spend per paid install at a 30% first-order rate, before promotions. The first-order event was missing a parameter in version 4.11; specifying it two weeks ahead meant version 4.12 reached about 70% of active users before the first campaign day.
Writing the plan before the campaign turned two hidden risks into dated tasks. The conversion value schema forced the team to agree that an order, not a browse, was the signal worth sending, and the release column showed that the activation event could not have been measured at all on the old version.
Measuring the plan: KPIs and cadence
The KPIs that matter for paid installs are first opens by campaign, cost per open, the activation rate within a fixed window, cost per activated user, and early retention read by install cohort. Each sits in the tool the plan made its owner. Health measures matter as much: the share of installs the partner could attribute, the deep-link success rate, the gap between store and partner counts, and the share of active users on a version that carries the current event spec. Read spend, opens and link health daily during a launch and weekly after it. Read activation and first-order rates weekly by install cohort, only once the window has closed. Read cohort retention monthly. Review the plan itself at every release that touches tracking, since that is when it most often goes out of date. The tutorial on checking app events on every release turns that review into a routine run on the release build itself, from clean test devices and the debug view to version adoption afterwards.
Where an app-install plan misleads
Aggregated postbacks hide which creative or audience drove the value, so the plan reports campaigns well and anything below them poorly. Modelled figures from networks and partners can fill gaps without saying how much is modelled. Organic installs rise when paid campaigns run, and none of the counters will credit that to the campaign. And a well-specified event can still be absent from most users for weeks, because the plan controls the release, not the update.
Common mistakes
- Writing the plan after the release. The campaign then measures whatever the old version recorded.
- Letting two tools report the same thing. Pick an owner for each question and ignore the duplicate.
- Accepting default conversion values. A schema nobody designed encodes nothing useful.
- Testing links only on installed devices. New users from a campaign take the deferred path.
- Chasing a perfect store match. The gap is structural; set a tolerance and watch for change.
- Ignoring version adoption. An event fix reaches users only as fast as they update.
