Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Mobile & App Marketing › Mobile Measurement

Setting an App's Lapse Threshold From Its Own Data

How to decide when an app user counts as lapsed: the core action, silences between actions, return rates by silence length, season checks and one written rule.
Tutorialadvanced15 min readUpdated

What a lapse threshold decides

The re-engagement guide defines a lapse as use that stops and rarely resumes unprompted, and says the right gap depends on how the app is used and comes from the app's own data. It names the method in one sentence: chart the chance of a return against how long each user has been silent, and put the threshold where that chance falls away. This tutorial is that sentence worked through, with the table, a query and the checks that stop the table misleading. Where a step goes beyond what the linked guides say, the text marks it as general practice.

One written rule has to serve every use below, because a separate number per team would make channels and reports disagree.

UseWhat the threshold does thereCovered in
Re-engagement triggerStarts a sequence for people silent that long, and, as general practice, a shorter line before it marks the people a message is most likely to matter forThe re-engagement guide
Paid re-engagement audienceDecides who is eligible, with the same definition as the owned channels, so paid and owned effort do not target different groupsThe paid section of the same guide
Churn and active-user reportsSets the inactive window behind "active" and "churned"; a window longer than one period means churn is counted only once the window has closedGrowth accounting, churn and retention metrics
Holdout assignmentFixes who is eligible, and when each person is assigned to a message or to the group that gets nothingWin-back offer test

A threshold set too short spends effort on people who were between sessions, and some of them were about to return, which flatters any campaign aimed at them. One set too long contacts people whose habit has already gone. A campaign report counts only the people who were contacted, so it cannot show who should have been contacted earlier, or not at all.

One limit applies throughout. The table below is built from what people did, and some of them did it after a message. If the app already sends re-engagement messages, past returns include returns those messages caused, so the curve overstates how readily silent users come back unaided. As general practice, build the table from a period before the programme, or from users who were held out, or record plainly that the history includes contact.

Before you start

The table starts from event-level data: one row for each time a user took the action that counts, with the date and a user identifier, as sent by the analytics SDK. The deep links and app analytics guide says app data reaches Google Analytics through the Firebase SDK and that teams needing more than the built-in reports export events to a warehouse. As general practice, treat this table as such an analysis: the built-in cohort and retention views group users by when they started, while this one groups silences by how long they have lasted, so build it in a warehouse. As general practice, a spreadsheet of exported user-days serves for a first look. The BigQuery export basics guide covers how GA4's export lands every event, with the app version and platform beside it, in tables the business owns. The cohort analysis tutorial builds the other table, retention by age, and is worth reading first if either is new.

As general practice, check what the user identifier does when someone reinstalls the app or moves to a new phone. Where it resets, a returning person looks like a new user and the old identifier looks lapsed from then on, which inflates the silences that never end.

Agree the core action before you start. The activation event tutorial chooses one from the app's own data, and the event taxonomy in the deep links guide holds the repeated core-value actions that follow it. Decide who signs off the rule. As general practice, the messaging owner, the analyst and whoever owns the product's core action sign it off, because each will use it.

Check the volume. Some lengths of silence will have few silences behind them, so follow the small numbers guide: a count beside every rate, a range round each figure and pooled periods wherever a cut is too thin.

The steps

  1. Write down the decisions the threshold will serve

    List every use from the table above that applies: a re-engagement trigger, a paid audience, a report's inactive line, holdout assignment. For each, note the owner and whether it needs one line or two. The re-engagement guide says users approaching the threshold are the ones a message is most likely to matter for, so a team that wants to act early needs, as general practice, a shorter at-risk line as well as the lapse line, and the same table reads both. Then write the question in one sentence: after how long a silence do people stop returning on their own? Everything below answers it.

  2. Choose the action that counts as being active

    Use the core action, not the session. The re-engagement guide warns against counting an app open to dismiss a notification as the behaviour that matters, and says that measuring the lapse against core actions finds people whose habit is breaking while they still open the app. Take the activation event the activation tutorial chose or, where the app has a repeated core-value action, that event from the taxonomy. Write its exact name, any parameter that qualifies it and the platforms it applies to. Then fix the unit. As general practice, count by calendar day in one written time zone, so that a day with at least one qualifying action is the unit, and finer units only for an app used in bursts within a day. The reconciliation tutorial explains why dates, time zones and platforms must line up before counts are compared.

  3. Decide who is in the data, and which dates to keep

    As general practice, include users who have taken the action at least once. Someone who installed and never did is an activation problem, which the diagnosis of installs not turning into active users separates, not a lapse. As general practice, carry the attributes you may split by later, taken as they were when the user started: platform, install cohort, install source and plan, and take out internal and test accounts. As general practice, keep out dates when the action's event was broken or double-counted: the measurement change log and the release checks tutorial show where a release moved an event, and a stretch with a broken event creates silences that never happened. As general practice, take enough history that the app's seasons, the parts of the year where use rises or falls, are all in the data, or silences that begin in a quiet season can be read as ordinary behaviour.

  4. Fix the follow-up window and the lengths to test

    For each length of silence the table asks one question: of the silences that lasted that long, how many ended within a further stretch of time? Decide that stretch, the follow-up window, before reading anything. As general practice, it should be long enough for the app's slower returners to show and short enough that the answer still guides who to contact now; a window as long as the whole history counts returns of any length. Then, as general practice, list the lengths to test, in steps that fit the app's rhythm: days for an app with a daily habit, weeks for a weekly pattern, longer for occasional use, and make the step equal to the window if you can, because each row's returns then equal the fall between two rows, which gives the check in step five. Write both down and hold them fixed for every cut, as the activation tutorial does with its window and retention measure, so that segments compare fairly.

  5. List the silences and count the returns

    As general practice, the unit is a silence: the run of days after a day with a qualifying action. It ends on the next qualifying day, or stays open if there is none. For each user list every silence, including the open one after the last action. For each length to test, count the silences that reached that many silent days, then count how many of those ended within the follow-up window after it. Keep open silences in the count, and leave out silences that began too recently for the whole window to have been observed. As general practice, use one cutoff for every row: the last complete day, minus the longest length, minus the window. The query below is general practice too, written in generic SQL; the date arithmetic and the generated list of lengths vary by warehouse.

    with qualifying_days as (
      select user_id, cast(event_time as date) as day
      from app_events
      where event_name = :core_action
        and cast(event_time as date) <= :last_complete_day
      group by user_id, cast(event_time as date)
    ),
    silences as (
      select user_id,
             day as last_action_day,
             lead(day) over (partition by user_id order by day) as next_action_day
      from qualifying_days
    ),
    lengths as (
      select k as silent_days from generate_series(:shortest, :longest, :step) as g(k)
    )
    select l.silent_days,
           count(*) as silences_that_reached_it,
           count(case when s.next_action_day <= s.last_action_day + l.silent_days + :follow_up_days
                    then s.last_action_day end) as returned_within_follow_up
    from silences s
    join lengths l
      on s.next_action_day is null
      or s.next_action_day > s.last_action_day + l.silent_days
    where s.last_action_day <= :cutoff_day
    group by l.silent_days
    order by l.silent_days;
    

    Three checks before trusting the output. The count of silences cannot rise from one length to the next. If the step equals the follow-up window, each row's returns equal its count minus the next row's count. And a run with the shortest length set to zero should return a count equal to the number of qualifying days up to the cutoff, since each qualifying day starts exactly one silence. The unit is a silence, not a user, so people with more qualifying days contribute many; the split by rhythm in step eight shows whether that distorts the answer.

  6. Read the return curve and find where it falls away

    Turn each row into a share: returns divided by silences that reached the length, with both counts beside it. Plot the share against the length of silence. As general practice, read the shape as a steep fall followed by a low stretch where the fall slows, the steep part being people who are between sessions and the low part people whose chance of coming back unaided has settled. Set the lapse threshold at the first length on or after the bend where the share has reached that low stretch and, as general practice, the at-risk line earlier on the steep part, where the share is still high but falling. Do not choose the single lowest point by eye from a thin cut: use the ranges and pooling the small numbers guide describes, and, as general practice, prefer a bend that holds across neighbouring rows. The worksheet below keeps every length on the same terms.

    Silent daysSilences that reached itReturned within the follow-upShare returnedNotes
    Each length testedCountCountReturns divided by silencesSample size, season, platform
  7. Check the result for survivorship and for recent silences

    Ask who is missing from the table. A view built only from silences that ended, the gaps between actions that people actually came back from, leaves out everyone who never returned, so it makes returns look sooner and more certain than they are. The survivorship bias framework names the general error of studying only what remains and its remedy, which is to bring the missing group back. That is why open silences stay in the count above. As general practice, build the returners-only version once and compare it with the full table, so the team sees how much it flatters. Then check the recent end of the data. The guide to comparing periods warns that figures for the latest days of a report are incomplete and says to compare at equal age, following each group for the same length of time. The same holds here: a silence still running at the end of the data looks like a non-return, which is why one cutoff removes silences too recent to judge and every silence is followed for the same window.

  8. Test the threshold by season, rhythm and platform

    Split the table and read each cut against the whole. By the season a silence began: use that rises and falls through the year can produce silences in the quiet season that end when the season returns, and a threshold read from the whole year could call those people lapsed. By how often people used the app before the silence: the sort the RFM segmentation framework calls recency and frequency, here read from core actions rather than purchases. By platform, since the app marketing troubleshooting section says to read iOS and Android separately, and by install source. If the curves are alike, keep one number. If they differ sharply and each group has enough silences to read, then, as general practice, write a rule per group. Where habits differ widely between individuals, a multiple of each person's own usual gap replaces a single number: the win-back diagnosis makes that case for purchases, with the gaps recalculated on a schedule.

  9. Test what the rule would do before anything sends

    As general practice, apply the threshold and the at-risk line to current users and count who crosses each in one recent week. Compare the volume with the old rule. Work out which channels still reach those people, using the table in the re-engagement guide, and sort them by recency, frequency and value so that the costlier touches go to people worth them. As general practice, read a sample of individual histories, to see that those flagged really went quiet and that nobody is flagged because an event stopped firing. Whether a message changes when people stop coming back unaided is the job of a holdout group, which the win-back offer test sets up for a lapsed audience.

  10. Write the rule and use it everywhere at once

    Write the definition in full: the qualifying action and its parameters, the length of silence for the at-risk line and for the lapse, how the clock starts (the last qualifying day), the time zone, the platforms or groups each rule covers, who is excluded and how off-season silences are treated, the evidence (period, counts, the table) and the follow-up window, with a date, an owner and a review date. As general practice, put it in the tracking plan beside the core action's event. Build the audiences from it, describe each in plain words and set membership durations deliberately, as the deep links guide advises. Give the same rule to the owned-channel trigger, the paid audience, the churn and active-user reports and the holdout assignment. The win-back and churn pages assume a definition like this exists; none of them can write it for the app.

  11. Revisit it when the app or its season changes

    A redesigned first session, a new core feature, a changed price or a release that renames the core event can move the curve. Rerun the table after any product change that touches the core action, and on a scheduled review date. When the threshold changes, change it in every use on one date, log that date in the change log, and expect trend breaks in churn and active-user counts, because the definition moved and the users did not.

A foraging log app finds its lapse threshold (illustrative)

Illustrative example — not a real result

Pondrissa Foraging replaces a seven-day rule

Pondrissa Foraging, an invented app for hobby foragers who log their outings and finds, sent "We miss you!" to anyone with no outing logged for seven days. Its analyst takes the core action to be a logged outing, not an app open, uses twenty-six months of outing events from about sixty-one thousand users, drops silences too recent to judge, and fixes a follow-up window of fourteen days with lengths tested every fourteen days. The table counts silences, not users.

Silent daysSilences that reached itReturned within fourteen more daysShare returned
742,00014,70035.0%
2127,3008,20030.0%
3519,1003,40017.8%
4915,7001,1007.0%
6314,6005003.4%
7714,100not readnot read

The checks hold: each row's returns equal its count minus the next row's. The 77-day row is the longest length tested and is there only so the 63-day row has a next row to check against, so its own share is not read. The share falls steeply, from 35.0% at seven days to 7.0% at forty-nine, and the fall slows after that, to 3.4% at sixty-three. Because the seven-day message went to everyone, the early rows include returns it caused, so the analyst reads them as an upper bound and puts the weight on where the fall slows: more than a third of the silences the old rule contacted ended within a fortnight, some of them because of the message. The team sets the at-risk line at thirty-five days, where 17.8% of silences still end within a fortnight but the share is falling fast, and the lapse at forty-nine, the first length on the bend where the fall slows. Splitting the forty-nine-day row by the season a silence began shows a seasonal pattern: 4,100 silences that began in the three winter months return at 12.0% in the following fortnight (490), against 5.3% for the 11,600 that began in other months (610). The 95 percent ranges, about 11.0% to 13.0% and 4.9% to 5.7%, do not overlap, a pattern that fits silences that begin in the quiet season ending when the season returns. Off-season silences are therefore held back from the lapse rule until the team has read a return curve for them of their own. Splitting by earlier use shows no difference worth a second rule, and the returners-only version of the table, built once, makes returns look both earlier and likelier than they are. Pondrissa writes the rule into its tracking plan, moves the push trigger, the paid audience and the churn report to it on one date and records the date in its change log. It keeps a holdout on the new trigger, because the table cannot say what a message adds.

Where the threshold goes next

The rule is an input, not an outcome. The re-engagement guide takes it into channel maps, reasons to return, paid campaigns, holdouts and the point to stop. The activation event tutorial defines the action that counts and this one defines the silence that counts as lapsed, and the measurement hub lists both beside the attribution and deep link guides. Retention by cohort, in the cohort analysis tutorial, reads the same actions by age since install, and the churn and retention metrics guide explains how the same inactive window shapes a churn rate. If many lapsed users share a reason for leaving, a threshold does not help; the re-engagement guide's advice is to fix the reason, then tell the people it affected.

Common mistakes

  • A borrowed number. A length taken from elsewhere has no link to this app's return curve. Read the length from the app's own return table.
  • Counting app opens as activity. Someone who opens the app to dismiss a notification has not returned to the product. Count the core action.
  • Leaving out the silences that never ended. A table built from the gaps people came back from makes returns look sooner and more certain than they are. Keep open silences in the count.
  • Reading silences too recent to judge. A silence still running at the end of the data looks like a non-return. Use one cutoff for every row.
  • Ignoring the season. Silences in a quiet season can end when the season returns. Split the table by the season a silence began.
  • One number for every habit. Where return curves differ sharply by rhythm, platform or season, a single threshold is wrong for some groups. Write a rule per group, or a multiple of each person's own gap.
  • Reading the table as proof a message works. Some people past the threshold return on their own. Use a holdout.
  • Changing the rule in one place. A channel on the old number and a report on the new one disagree until both use the same rule. Change every use on one date and log it.

Frequently asked questions

Why count silences instead of users?

One person can go quiet several times, and each gap ends in its own way. A count of users would have to pick one gap per person, and the obvious choices each hide something: the first gap ignores how the person behaves later, and the latest gap may still be open. Counting silences keeps every gap, and open silences stay in the count, so people who never came back are not dropped. The cost is that people with more qualifying days contribute more silences, which is why the table is split by rhythm before the number is settled.

Does the threshold prove that a message will bring lapsed users back?

No. The table describes silences, whether or not they ended in a return and with or without any contact, so it says when returning on their own becomes unlikely, not what a message adds. Some people past the threshold will still return unprompted. To learn what a message, an offer or a paid audience adds, randomly assign a share of eligible users to receive nothing and compare their later returns with those of the users who were contacted.

Should every user share one lapse threshold?

Only if the return curves are alike. Split the table by the season a silence began, by how often people used the app before, by platform and by install source. Where the curves differ sharply and each group has enough silences to read, a rule per group is the general practice; where they do not, one number is easier to run, test and report. Either way the rule is written once and used by every channel and report.

Terms used on this page

Frameworks

Sources

  1. Firebase documentation — Google
  2. Google Analytics Help — Google Analytics Help

Related