Get in Touch

SolutionsProduct Retention

Churn is visible weeks before the cancel. Most brands only look after.

Most churn is decided in the first 30 days, and the rest shows up in behavior first. We build the cohort economics that show which habits keep customers and which quietly lose them, then act per customer, including not touching them at all.

Where engagements start

Where retention usually stands when we arrive.

"Is retention getting better or worse? Depends who you ask."
The blended rate moves with acquisition and mix as much as with behavior, so it usually ends with retention blaming the quality of new customers while marketing points back, each armed with a different dataset. Cohorts at the same age, on one set of definitions, settle it, and they are the first thing we build.
The promo calendar is the retention plan
It works, by its own numbers. The read that changes the plan is how much of each promo went to people who would have bought anyway, because that share is margin, not retention.
Lifecycle flows that have run untouched for a year
By opens and last-click they all work. A small holdout says which ones add orders that would not have happened anyway, and it is usually a mix worth knowing.
Churn arrives as a surprise
It rarely is one. Skips, pauses, order cadence and delivery issues move weeks before the cancel, if the events exist to see them, which is what the tracking plan below is for.
Onboarding exists, activation is undefined
A completed funnel step is not activation. The early behaviors that matter are the ones that predict a return, which is a modeling question, and every business's answer is different.

None of this needs another platform. It needs cohort economics, holdouts, and a read on what actually drives your curve. Here is what that found at one business.

Findings from the work

What drove retention at a $400M subscription business.

We ran driver discovery there: churn and order prediction models plus cohort reads across the whole base, then tests on what surfaced. What actually moved retention looked nothing like the promo calendar. Six of the findings:

First 30 days~Half of all churn happened here, mostly before a second order.
~3x
The first automated order

Churn among customers whose first auto-order landed before they trusted the service.

Default suppressed, heads-up instead. Biggest early win that year, and it cost nothing.
17→40%
Preferences set at onboarding

Completing setup predicted better week-4 retention and higher revenue per customer.

Onboarding rebuilt around finishing it.
Never
Coming back at all

A large share of early churners never returned to browse after signup.

The first week became the program, read like a funnel.
Month 2 onwardDifferent drivers take over, and none of them is a coupon.
40/60
Repeat-to-discovery mix

Best retainers repeated ~40% and discovered ~60%. Either extreme retained worse.

Merchandising tuned to protect the mix, not maximize novelty.
No.1
Skip & pause behavior

Skipping far ahead of the deadline was the strongest churn signal, weeks before a cancel.

An early-warning list the lifecycle program acts on.
Top 5
Order experience quality

Share of issue-free orders ranked among the top churn predictors, with a lag of weeks.

Ops metrics joined to retention economics.
01
Found in behavior
Cohort reads and ML feature importance surface the candidates. At this stage they are correlations, and we say so.
02
Proven with a test
A/B or holdout, read on the segment the change actually touches with behavioral evidence, not just the end KPI.
03
Scaled with models
Churn risk and next-action models score every customer weekly, so the finding keeps paying instead of becoming a slide.

The foundation nobody sees

Instrumentation that can see retention coming.

Most stacks record the order and nothing around it. The reads above live in the events between orders: skips, preferences, saves, holdouts, delivery outcomes, designed once with engineering.

Retention tracking plan · excerptEvery event QA'd with engineering
Order Completed
Core
order_number_for_customerdays_since_last_orderfirst_orderitemsdiscount_codemargindelivery_promise_date
Product / Catalog Viewed
Browse
days_since_last_visitnew_items_seencategorysource
Preferences Set
Activation
completedweek_of_lifefields_filledchannel
Skip / Pause
Subscription mechanics
days_before_cutoffreason_selectedorders_streak_beforeplan_size
Lifecycle Message
CRM
holdout_groupaction_chosenoffer_valuecampaignclicked
Rating & Review Submitted
Product feedback
ratingproduct_idorder_idrepeat_intent
Support Contact
CX
reasonresolutioncsatorder_id
Cancellation Flow
Churn
reasontenure_weekssaves_shownsave_accepted
Delivery Outcome
Ops
on_timeissue_typeresolutionorder_id
Half the reads above cannot be computed on a standard setup. days_before_cutoff and holdout_group are where the retention findings live, and every new feature ships with its events built and QA'd, so coverage does not decay.

Excerpt of the shape. The full plan carries definitions and QA status per event and property.

Then ask it anything

The same data, answering in plain language.

On top of the instrumentation sits a governed AI layer: your definitions, your caveats, encoded once, so anyone on the team can ask retention questions and trust what comes back.

Head of Retention

Q1 cohorts look worse versus last year. Is that real, or is it the promo mix?

Governed layer · live

At the same age, Q1 cohorts repeat ~1pt below last year. About 70% of the gap is mix: a larger share of first orders came through the deep-promo channel, which repeats lower. Same-channel cohorts are roughly flat, so behavior has not worsened.

Using: cohort repeat rate v3 (finance-approved) · channel mix definitions last reviewed with the growth team

How the governed AI layer works →
The playbook

Cohort economics first. Then the program runs on them.

You do not need all of it, and almost nobody starts with more than two. Everything lives in your repo and your warehouse, so it is yours whether or not we are around.

01
Cohort & LTV foundations
Repeat, revenue and margin by cohort at the same ages, defined once and reused everywhere retention gets discussed.
02
Retention driver discovery
Models and cohort reads that rank which behaviors separate keepers from churners, plus the user research that says why.
03
Per-customer next action
AI that picks each customer's next touch, offer, content or nothing, optimizing toward the same revenue number finance reports.
04
Discount governance
Who gets an offer, who would have bought anyway, and what the discount actually bought. The cheapest margin win most brands have.
05
Lifecycle with holdouts
Welcome, winback and replenishment flows measured against people you deliberately did not touch, so programs earn their place.
06
Onboarding & second order
The window where repeat is won or lost, instrumented: what first-order experience predicts a second, and what accelerates it.
07
Products that create repeat
Which first purchase predicts the customers worth keeping, so merchandising and acquisition lead with it.
08
Subscription mechanics
Pause, skip, plan changes and dunning read on cohort economics, so saves target value instead of noise.
Acting on it

The right next action for each customer, including none.

A promo calendar treats everyone the same, which is exactly how margin leaks. Decided per customer, the biggest single win is usually the offer you do not send, and holdouts keep the whole thing honest.

This is the pattern behind the +10% revenue per user result at the business above, and it only works on top of the cohort economics.

See how the model decides, in the full study →
Next-action decisions · one morning
41%
No touchHigh likelihood to buy anyway. The discount stays in your margin.
32%
Content, not discountNew customers inside the onboarding window get education, not a coupon habit.
22%
OfferAt risk and worth keeping, sized by projected LTV.
5%
HoldoutDeliberately untouched, so every program above proves itself against them.
Illustrative split. Every decision is logged with its reason, so the program can be taken apart by a skeptic.

Get in touch.

Tell us what you are trying to figure out. If we can help, we will say how. If we are not the right fit, we will say that too, and point you somewhere better.

01
Free work first
We start by doing a real piece of work on your data, free, so the first thing you judge is output, not a pitch.
02
First build, fixed scope
A trusted core metric layer in weeks, not quarters, at a fixed price. Small enough to be safe, real enough to matter.
03
Scale when it earns it
Month to month from there, from a few thousand a month. You own every model, dashboard and definition.
Book a callhello@meridiangrowth.ai