Push vs in-app: what each is for
Push and in-app messages are often bought as one product and planned as one channel, but they do opposite jobs. A push notification is an interruption delivered through the operating system to someone who is doing something else; its job is to earn a return visit. An in-app message is shown inside a session the person chose to start; its job is to help them get more out of the visit they are already making. Confusing the two produces push notifications that explain features nobody is using and in-app banners that try to sell to people in the middle of a task. Which in-app format suits which job, and how to place, word and cap those messages, is the subject of the guide to designing in-app messages.
| Row label | Push notification | In-app message |
|---|---|---|
| Where it appears | Lock screen, notification centre, banner over other apps | Inside the app: modal, banner, tooltip, card, inbox |
| Who it reaches | Installed users who have allowed notifications | Anyone currently in a session |
| Permission | Operating-system opt-in, revocable at any time | None beyond the session itself |
| Main job | Bring people back for something timely or personal | Guide, explain, offer at the moment of use |
| Content length | A title and a line or two, plus optional image | As much as the format allows, though shorter still wins |
| Main risk | Disabling notifications or uninstalling | Annoyance, dismissal blindness, blocked tasks |
| Typical trigger | Something changed outside the app, or a lapse in activity | Something the person just did, or a screen they reached |
In the AARRR model, in-app messages do most of their work on activation (getting a new user to the moment the app makes sense) and push does most of its work on retention (bringing them back after they have drifted). Both feed revenue and referral, but neither should be planned from those stages first. The wider programme they sit inside, from onboarding to win-back, is the subject of the lifecycle marketing guide.
Earning the permission
On iOS an app must ask before it can send notifications, and the system prompt can be shown only once; if the person declines, the only way back is through the device settings. Recent Android versions have moved to the same model, with an explicit runtime permission for notifications. The permission is therefore the single most valuable asset in the channel, and asking for it on first launch, before the person knows what the app does, spends it at the worst possible moment. The share of people asked who allow it, the push opt-in rateShare of app users shown the notification permission prompt who allowed notifications., caps every push campaign that follows.
The pattern that works is a soft prompt followed by the system prompt. The soft prompt is an ordinary in-app screen or card that explains what notifications will do for this person, shown at a moment when that value is obvious: after they place an order they will want to track, save a search they want alerts on, or book something they could forget. If they agree, show the system prompt; if they decline, the system prompt has not been used and can be offered again later. Apple's App Store Review Guidelines also say that push notifications must not be required for an app to function and must not be used for marketing unless the person has opted in to receive it, so the soft prompt should be honest about promotional messages if the app intends to send them.
Two further practices protect the permission once granted. Offer notification preferences inside the app, by topic, so a person annoyed by one kind of message can turn that kind off rather than turning off everything. The tutorial on setting up notification preferences by topic works through the inventory, defaults and settings screen that make this possible. And track the enabled rate over time, not just at install: a falling share of users with notifications on is the earliest sign that the channel is being overused.
To check the timing on your own app rather than assume it, the push permission prompt test compares a first-launch prompt with an explained, delayed one on opt-in among all new installs, with early retention as the guardrail.
When the opt-in rate is already low and the cause is not yet known, the diagnosis of a low push notification opt-in rate checks the measurement first and then works through the prompt's timing, the soft prompt, its wording and the topic choices.
Message types and triggers
Messages fall into a few types, and each has a natural trigger. Planning by type keeps the programme balanced; planning by campaign calendar tends to fill it with broadcasts.
| Type | Example | Trigger | Channel |
|---|---|---|---|
| Transactional | Order shipped, payment failed, booking confirmed | A system event the person caused | Push, with in-app record |
| Activity | Someone replied, your item is back in stock, a price dropped | An event involving the person's own content or saved items | Push |
| Onboarding | Finish setting up, try the feature that matters | Steps not yet completed after install | In-app first, push if they lapse |
| Feature discovery | A tip for the screen they are on | Reaching a screen or completing a task | In-app |
| Re-engagement | Something new since your last visit | A period without a session | Push |
| Promotional | A sale, an offer, a new collection | The calendar or a segment rule | Push sparingly, in-app inbox |
Triggered messages, those caused by something the person did or something that changed for them, almost always outperform broadcasts on both response and tolerance, because they are relevant by construction. Broadcasts still have a place for genuine news, but they should be the minority, segmented rather than sent to everyone, and counted against the same frequency cap as everything else.
The onboarding sequence deserves particular care, because it is the app's equivalent of an email welcome series and follows the same logic: one goal per message, the next most useful step, stop when the person has done it. The welcome series guide explains that logic for email; in an app, the first steps usually belong in-app, with push reserved for people who leave before finishing. Where new users also receive email, the tutorial on building onboarding emails for a software product fits the email side around the in-app messages, so that the same nudge never arrives twice. The tutorial on onboarding messages for new app users builds the app side of that sequence, triggered by the step a new user has not taken and stopped by the one they have. Firebase's documentation describes both channels (Cloud Messaging for push and In-App Messaging for in-session campaigns) with event-based triggers, and most dedicated engagement platforms work the same way.
Timing, frequency caps and quiet hours
Timing has two parts: when in the day, and when in the person's relationship with the app. Time of day matters less than relevance, but a notification that arrives at night in the recipient's own time zone does damage regardless of content, so quiet hoursThe hours when promotional texts, WhatsApp messages and push notifications are held for the recipient's daytime, set by law in places and courtesy elsewhere. should be a system rule rather than a campaign setting. Schedule in local time, hold non-urgent messages that fall inside quiet hours until the morning, and let transactional messagesAn email triggered by a person's own action or account, such as a receipt, password reset or shipping notice, rather than a marketing send. through because the person is expecting them.
FrequencyAverage times a person saw an ad. needs a cap per person, per period, across every campaign and every team that can send. Without a shared cap, each campaign looks reasonable on its own and the person receives all of them. Most engagement platforms let a team set global caps and priority rules, so that when several messages qualify the most important one is sent and the rest wait or drop. Transactional messages should sit outside the cap; promotional messages should sit at the bottom of the priority order.
The relationship timing matters as much. A new user in their first week can tolerate more guidance than a long- standing user, who needs very little. A lapsed user needs one well-chosen reason to return, not a daily reminder. Intelligent delivery features, where the platform chooses the send time per person from past opens, can help at the margin, but they do not rescue a message nobody wanted. Watch the leading indicators of overuse together: notification disables, topic opt-outs, uninstalls in the days after a send and falling enabled rates.
Deep links and landing in the right place
A push notification that opens the app's home screen throws away most of its value. The person tapped because the message promised something specific: their order, the reply, the item back in stock. Every notification and every in-app call to action should carry a deep linkA link that opens a specific screen inside a mobile app rather than its home screen. to that exact screen, with the right item loaded and any filters applied, so the tap continues the thought the message started. Testing each link on every platform is one line of the mobile message send QA checklist, the gate to tick before a push or in-app message is sent.
Deep links fail in predictable ways. The person is logged out and the link drops them on a login screen with no memory of where they were going; the item has sold out or been removed; the screen has been renamed in a release and the old route no longer resolves. Each case needs a designed fallback: return to the intended screen after login, show a sensible alternative when the item is gone, and test links against every live app version, not just the newest. Attach campaign parameters to the link so the analytics can tell which message produced the session.
The mechanics of link types, deferred links for people who do not yet have the app, and the events that record what happened after the tap belong to the measurement section. The deep links and app analytics guide covers how to build the links and the event taxonomy that makes message performance readable.
Testing and measuring messages
The metrics every platform reports by default (sends, deliveries, opens, click-through or CTRShare of impressions that result in a click.) describe the message, not its effect. A notification that says "your order has shipped" will be opened by many people who would have opened the app anyway. The question that matters is incremental: did the people who received the message do more of the target action than comparable people who did not?
Answer it with a holdout. For any recurring message or journey, keep a random share of eligible users who receive nothing, and compare the target behaviour between the groups over a sensible window: sessions, conversion (CVRShare of visits or clicks that convert.), retention at a set point, revenue. A difference between groups is the message's effect; a difference that is negligible means the message is noise, however good its open rate. Keep the holdout in place for long enough to see whether a lift persists or simply pulls activity forward.
Content tests work inside that structure. An A/B TestControlled comparison of two variants. on copy, timing or image tells you which version of a message works better; the holdout tells you whether sending it works at all. Firebase and most engagement platforms support both. Measure the costs as well as the gains: opt-outs, disables and uninstalls by message and by group, because a message that adds a little activity and costs reach has to be judged on both.
Corrie Fitness tests its daily reminder
Corrie Fitness, an invented workout app, sends a daily 7am "time to train" push to all 60,000 users with notifications on. It runs a four-week holdout: 54,000 users keep the reminder and 6,000 receive nothing. The reminder's open rate is 9%. Workouts per active user are 2.31 a week with the reminder and 2.24 without, a difference within noise. Notification disables run at 1.8% a week in the reminder group and 0.4% in the holdout. Corrie replaces the daily broadcast with a triggered push sent only after three days without a workout, at the user's usual training hour. In a second test that version lifts weekly workouts to 2.46 against 2.25 for its holdout, with disables at 0.5%.
Because opt-in sets the size of the audience, RACE for app installs makes permission the first Engage objective before any re-engagement campaign is budgeted.
Common mistakes
- Asking for permission on first launch. The single system prompt is spent before the person knows what notifications would do for them. Use a soft prompt at a moment of obvious value.
- No cross-campaign cap. Each team's messages look modest on their own and overwhelming together. Set one cap per person per period and a priority order.
- Broadcasts as the default. Messages sent to everyone on a calendar are tolerated least. Make triggered messages the majority and segment the rest.
- Push that opens the home screen. Every tap should land on the screen the message promised, with a fallback for logged-out users and missing items.
- In-app messages that block tasks. Full-screen modals on every open teach people to dismiss without reading. Show in-app messages at natural pauses and trigger them from behaviour.
- Judging messages on opens. Opens measure attention, not effect. Use holdouts to measure what the message changed, and count disables and uninstalls as costs.
- One switch for everything. Without topic preferences, a person annoyed by promotions turns off order updates too. Offer granular settings inside the app.
- Ignoring time zones. A well-written message delivered at night costs more goodwill than it earns. Enforce quiet hours in local time as a system rule.
