Retention Cohorts: The Only Growth Metric That Survives a Budget Cut

Every brand looks healthy when the ad account is open. Spend goes up, orders go up, the dashboard is green, and everyone agrees growth is working. Then the budget tightens. A funding round slips, a festive quarter underperforms, a board asks for profitability instead of topline. The spend comes down. And within a few weeks you learn the truth about your business: how much of that revenue was yours, and how much you were renting from Meta and Google.

The metric that answers this is not blended ROAS. It is not topline growth. It is cohort retention. Repeat-purchase cohorts are the one growth number that does not collapse the moment you stop paying for traffic. Everything else is borrowed.

Most growth is rented, and the lease is short

Here is the uncomfortable framing we bring to founders. If your revenue is flat to your ad spend, you do not have a brand. You have a paid-traffic arbitrage that happens to sell a product. The day the arbitrage stops being profitable, the revenue stops too. You were renting growth, and the landlord just raised the rent.

This is not hypothetical. Indian D2C has lived through it twice in recent years. CACs that looked sustainable at one funding climate became indefensible at the next. The brands that survived were not the ones with the cleverest creative. They were the ones whose existing customers kept buying without being re-acquired. That is cohort retention, and it is the load-bearing wall of any business that intends to outlast a cheap-capital cycle.

Acquisition is how you meet a customer. Retention is whether you have a business. Brands confuse the two until the budget gets cut, and then the difference is the only thing that matters.

What a cohort actually tells you

A cohort is simply a group of customers who first bought in the same period, tracked forward over time. You watch what share of each month’s first-time buyers come back to purchase again in month two, month three, month six. Plot a few of these side by side and the shape of your business becomes impossible to hide from.

The single most useful read is whether your retention curves are flattening or decaying to zero. A curve that drops fast and keeps dropping means every rupee of growth has to come from new acquisition forever. A curve that drops and then flattens means you have a base of customers who stay. That flat tail is the asset. It compounds. It is the part of the business a budget cut cannot touch.

Month-two repeat rate. The earliest honest signal of whether the product and first experience earned a second order.
Curve shape over six months. Decay to zero versus a flattening tail tells you if you own customers or merely visited them.
Cohort-over-cohort drift. Are newer cohorts retaining better or worse than older ones. Worse means your acquisition is buying lower-quality demand even as the dashboard looks fine.
Category-honest cadence. A supplements brand should see reorders in weeks. A mattress brand will not. Judge the curve against a realistic repurchase interval, not a generic benchmark.

Why blended metrics hide the rot

The reason most brands do not see the cliff coming is that their headline metrics blend new and returning revenue into one number. Blended ROAS looks fine because returning customers, who cost almost nothing to convert, subsidise the true cost of acquiring new ones. The average looks healthy while the underlying acquisition economics quietly rot.

This is the same disease we describe in our piece on why blended CAC lies to you. A blended number is an average that hides its own composition. When you separate first-purchase economics from repeat economics, you usually find that new-customer acquisition has been unprofitable for months, propped up by a loyal base you have been taking for granted. Cut the spend and the subsidy disappears with it.

The LTV trap

Brands love to quote a lifetime value number to justify a high CAC. The problem is that most LTV figures are projections built on the assumption that retention holds. If the cohort curve is decaying, your real LTV is a fraction of the modelled one, and you have been overpaying for acquisition against a fantasy. LTV is an output of retention, not an input you get to assume. Measure the cohort first. Let it tell you what a customer is worth. Then decide what you can afford to pay to acquire one.

Retention is built before the budget gets cut, not after

The cruel part is that you cannot fix retention reactively. When the budget is already being slashed, you are out of time to build the habits, the channels, and the trust that bring customers back. Retention is infrastructure. It has to be laid down while acquisition is still flowing, which is exactly when most brands ignore it because the topline looks great.

This is why we treat Retention & Lifecycle Marketing as a load-bearing function from day one, not a phase you graduate into. The work is unglamorous and it compounds. It starts at launch, which is why we fold it into the operating plan we describe in the first 90 days of launching a D2C brand in India, rather than bolting it on once the ad costs hurt.

In the Indian context, the highest-leverage retention channel is usually not email. It is WhatsApp, used with discipline rather than as a broadcast hose. Done right, it earns repeat orders at a fraction of re-acquisition cost, which is precisely what props up a cohort curve when paid traffic dries up. We lay out the operator approach in using WhatsApp as a retention channel for Indian eCommerce. The point is not the channel. The point is owning a direct line to customers you have already paid to acquire.

How we read a cohort table as operators

When we audit a brand, the cohort table is the first thing we pull and the last thing we trust the founder’s gut over. Here is the operator lens we apply.

Strip out returning revenue and look at new-customer acquisition on its own. If it cannot stand alone, the business has a hidden dependency on a base it is not protecting.
Read the curve against the category’s real repurchase cadence, not a borrowed benchmark from a different vertical.
Watch newer cohorts against older ones. Deteriorating cohorts are an early warning that scaling spend is buying worse customers, long before ROAS shows it.
Pressure-test every LTV claim against the actual flattening of the tail, not the slope someone wishes were true.
Decide acquisition budgets off proven repeat behaviour, so that a budget cut trims fat and not muscle.

None of this requires exotic tooling. It requires the willingness to look at the one table that tells you whether you are building equity or burning cash with extra steps.

What changed recently

The market has caught up to the argument. The cheap-capital playbook of buying growth and worrying about retention later is being retired in public, not just in operator slide decks.

Inc42 now frames the current phase as D2C 3.0, where growth is described as becoming less dependent on performance marketing and more driven by retention, repeat-purchase behaviour, operational efficiency and owned consumer relationships. The same piece names the structural killers behind failed brands as low repeat percentage, weak differentiation, high customer acquisition costs and thin unit economics. That is the cohort argument restated as a post-mortem. The brands that did not measure the curve are the cautionary tales.

The evidence is now showing up in the numbers, not just the narrative. YourStory reported that some D2C brands saw close to a doubling of festival-season sales in 2024 over 2023 and a further roughly 50 percent rise in 2025, and noted that brands which built owned audiences are benefiting from lower acquisition costs and better retention while marketplace-first brands keep chasing profitability. Owned audience is just another name for a cohort you can reach without paying for the click again.

Even quick commerce, the most acquisition-heavy corner of Indian retail, is making the same turn. Reporting in Business Standard describes the sector shifting from speed and cash burn toward density, advertising yield and retention rates as the levers that matter, with Zepto closing the order gap on Blinkit even as profitability stays out of reach. When the players burning the most money start optimising for retention instead of raw growth, the question for a smaller D2C brand answers itself. If platforms with infinite balance sheets cannot outrun weak repeat behaviour, neither can you.

The metric that survives the cut

Budgets always tighten eventually. Capital cycles turn, quarters miss, priorities shift from growth to profit. When that day comes, every brand finds out what it actually built. The ones who treated acquisition as the whole game watch revenue fall in lockstep with spend. The ones who built retention watch their cohorts keep paying.

That is why cohort retention is the only growth metric we treat as truly load-bearing. It is the number that holds when the ad account goes quiet. Build it early, measure it honestly, and let it govern what you are willing to pay for the next customer. Do that and a budget cut becomes a pruning. Ignore it and a budget cut becomes an ending.

Customer LTV on Marketplaces: Estimating What You Cannot Directly See

Lifetime value is the number that decides how much you are allowed to spend to win a customer. On your own store you can calculate it almost cleanly. You know who bought, how often they came back, and what each order was worth. On Amazon or Flipkart you know none of that at the person level. The platform owns the buyer and hands you aggregates. So most marketplace brands set LTV aside as a D2C luxury, cap their acquisition cost at the first order, and quietly underspend on the categories that would have paid them back twice over.

That is a mistake, and it is the expensive kind, because it is invisible. You never see the second order you failed to fund. The honest position is not that marketplace LTV is unknowable. It is that you cannot measure it directly, so you have to estimate it deliberately, from signals the platform leaks, and treat the estimate as a range you can defend rather than a number you can prove.

Why LTV is worth estimating even when it is fuzzy

The point of LTV was never precision. It was a spending ceiling. If a buyer is worth eight hundred rupees over their life with you, you can rationally pay more to acquire them than if they are worth two hundred. That decision does not need a perfect number. It needs a defensible band and the discipline to act on it.

The brands that refuse to estimate LTV do not avoid the decision. They make it implicitly, by capping acquisition cost at the first-order margin. That cap is a hidden LTV assumption of exactly one order. For a consumable that reorders every month, that assumption is wrong by a wide margin, and the competitor who modelled it correctly will outbid you on every keyword that matters.

Refusing to estimate LTV is not caution. It is choosing the worst possible estimate, one order, and pretending you did not choose.

The signals that proxy lifetime value

You will never get a customer ID you can build a clean LTV model on. You will get enough to build a coarse one. Three signals carry most of the weight for marketplace brands in India.

Subscribe-and-save enrolment and churn. If you sell anything repeatable, this is the closest thing to a real retention curve a marketplace gives you. Enrolments and how many survive three and six months later are a near-direct read on how long a buyer stays. Treat the subscribe cohort as your highest-confidence LTV input.
Repeat-purchase rate at the SKU level. Amazon and, less generously, Flipkart report repeat behaviour in aggregate. The repeat rate is the probability of a next order. Stack a few of those and you have an expected order count, which is the spine of any lifetime-value estimate.
New-to-brand share against total orders. This tells you how much of each period is fresh acquisition versus returning demand. A buyer base that keeps returning is a base with real lifetime value. One that is all new-to-brand every month has none, no matter how good the topline looks.

None of these names a single customer. Together they describe the curve, and the curve is what LTV actually is. The work of turning these into a number is cohort work, which is why the two disciplines are inseparable. We lay out the proxy method in full in building cohorts without first-party data, and an LTV estimate is the natural output of those cohorts rather than a separate exercise.

A method you can defend in a meeting

Keep the model crude on purpose. A crude model you can explain beats a precise one nobody trusts. The estimate has three inputs and one honest caveat.

Expected number of orders. Start from the SKU repeat-purchase rate. If roughly a fifth of buyers place a second order and a smaller share a third, you can build an expected-orders figure without ever knowing who those buyers were. For subscribe SKUs, use the survival curve instead, it is cleaner.
Contribution per order. Use contribution margin, not revenue. Revenue-based LTV flatters you into overspending. Net of fees, ad cost, returns, and shipping is the only version that maps to what you can actually afford. This is the same discipline we apply when reading profitability one SKU at a time, and LTV inherits its honesty from that per-SKU margin work.
A decay assumption. Buyers churn. Apply a falling probability to each successive order rather than assuming a flat repeat rate forever. A simple decay is enough. The goal is a band, low and high, not a single confident point.

Multiply expected orders by contribution per order, discount the later orders, and you have a lifetime-value range per SKU or per category. State it as a range every time. The honesty of the range is what makes leadership trust the rest of the analysis.

Where the estimate breaks, say so first

The model has known weaknesses, and naming them yourself is what keeps it credible. Repeat rates from a promo-heavy period overstate loyalty, because deep-discount buyers repeat worse than full-price ones. A short observation window understates LTV for slow-reorder categories. And aggregate repeat data cannot separate one devoted buyer placing five orders from five buyers placing one each. Flag each of these in the room before someone else does, and the estimate survives the scrutiny it deserves.

What the estimate actually changes

An LTV band is useless if it does not move money. It should change three decisions directly.

Your acquisition ceiling. A SKU with a strong subscribe curve and a rising repeat rate can justify a first-order acquisition cost that looks reckless in isolation, because the second and third orders repay it. A SKU that never repeats has to win on the first order or be cut.
Where you scale spend. Two SKUs with identical first-order economics are not equal businesses if one has triple the expected order count. LTV is what separates them, and without it you fund both as though they were the same.
Who you chase back. The cohorts with the highest estimated value are the audiences worth re-engaging. That points budget straight at the demand most likely to return, which is the entire logic of retargeting marketplace shoppers when you do not own their data. You spend re-engagement money where the lifetime value is, not evenly.

Notice that every one of these is an acquisition decision. That is the point of the angle. You estimate LTV not to admire it but to know how aggressively you are allowed to spend at the top of the funnel. Subscribe-and-save and repeat signals proxy that value well enough to set the ceiling, and a deliberately rough estimate beats the silent assumption of one order every time.

What changed recently

The platforms have started handing you a cleaner subscribe signal than they used to, and the quick-commerce players have turned subscription into the explicit retention lever. That is good news for anyone estimating LTV, because it makes the highest-confidence input bigger and more legible.

The clearest proof is the subscription membership data itself. When Zepto piloted its paid membership, subscribers spent more than 30 percent more on the app and their monthly retention improved by roughly 10 percent, with the pass crossing a million members inside a week of launch, per YourStory. Read that as an LTV statement, not a marketing one. A membership cohort that spends a third more and churns less is a directly observable retention curve, the exact thing marketplace aggregates usually deny you. If you sell through a platform with a membership tier, the subscriber cohort is now your highest-confidence band, and you should be modelling it separately from walk-up demand.

The strategic backdrop reinforces the same point. Inc42’s 2026 quick-commerce outlook argues there is still Inc42 virtually no loyalty in the category, with shoppers keeping several apps and switching on availability and discount, and it expects growth to come from getting existing customers to order more often rather than from new cities, with subscriptions and private labels as the main stickiness levers. For a brand, that is a direct instruction. The platforms are now optimising for order frequency, which means the repeat and subscribe signals you build LTV on are about to carry more of the platform’s own attention, and the brands that already have a defensible LTV band will set acquisition ceilings the switchers cannot.

Make the number live somewhere leadership will look

An LTV band that lives in one analyst’s spreadsheet changes nothing. It has to sit beside acquisition cost and per-SKU margin in the same view, so the spending decision is obvious at a glance. The relationship between what a buyer costs and what a buyer is worth is the single most important comparison on a marketplace P&L, and it deserves a permanent line, not a quarterly slide. Getting it there without drowning the reader is its own craft, the one we mean by a reporting dashboard leadership will actually read.

Update the estimate as the cohorts mature. Early repeat data is noisy, and an LTV band built on two months of signal should widen, not pretend to certainty. As subscribe curves lengthen and repeat rates stabilise, the band tightens, and your acquisition ceiling earns more confidence with it. On quick commerce specifically, the same logic governs unit economics once platform fees are netted out, which is why we treat LTV and quick-commerce unit economics after platform fees as two reads on the same P&L.

The short version

You cannot see a marketplace customer twice. You can still estimate what they are worth, from subscribe survival, repeat-purchase rates, and new-to-brand share, assembled into a contribution-based lifetime band with an honest decay. The estimate will be a range, and that is correct. It is still enough to set an acquisition ceiling that beats the silent one-order assumption most brands spend under. And as platforms lean harder into membership and frequency, the subscribe signal at the heart of that band is only getting stronger.

Our Analytics & Reporting work exists to turn those scattered platform signals into an LTV band a brand can spend against, and our Marketplace Performance teams use it to decide where acquisition money actually compounds. The customer is hidden. Their value is not. Estimate it, defend the range, and spend like you mean it.

Cohort Analysis for Marketplace Brands Without First-Party Data

Every retention playbook written for D2C assumes one thing you do not have on a marketplace. The customer. On your own store you have an email, a phone number, an order history tied to a person. On Amazon or Flipkart you have a settlement report and a wall. The platform owns the buyer, guards the identity, and hands you aggregates. So most marketplace brands quietly conclude that cohort analysis is a luxury for people with first-party data, and they stop looking.

That conclusion is wrong, and it is expensive. You cannot run a textbook cohort table keyed to individual customers. But the marketplace leaks enough repeat-behaviour signal that you can build cohorts that are directionally true and good enough to change what you spend, what you stock, and what you launch. The trick is to stop trying to reconstruct the customer and start reading the signals the platform cannot hide.

What you are actually missing, and what you are not

Be precise about the gap. The thing you lose on a marketplace is identity resolution. You cannot reliably say buyer 4471 purchased in January and again in April. Amazon’s brand tooling gives you a new-to-brand flag and some repeat-purchase aggregates, Flipkart gives you less, quick commerce gives you almost nothing at the person level. None of it is a clean customer ID you can build a classic cohort grid on.

What you keep is more than people assume. You keep the new-to-brand percentage on advertised sales. You keep subscribe-and-save enrolment and churn if you sell consumables. You keep total repeat-purchase rate at the account or SKU level where the platform reports it. You keep your own units-per-order and the gap between gross units sold and unique-ish demand. And you keep time. Every one of those is a retention signal. Cohort analysis without first-party data is the discipline of assembling those signals into a defensible view of whether buyers come back, even when you can never name a single one.

You are not reconstructing the customer. You are reconstructing the curve. The curve is what actually drives the decision.

Build cohorts on proxies, not people

If you cannot cohort by customer, cohort by the next best thing. The most useful unit is the acquisition window. Group everything by the month a buyer most plausibly entered the brand, then watch how the brand’s behaviour evolves against that window. You will not have per-person retention, but you will have a brand-level repeat signal that moves with the cohort.

Three proxy cohorts do most of the work for marketplace brands in India.

Launch-month cohorts. Tag the month a SKU or variant went live. Track repeat-purchase rate and subscribe enrolment for that SKU over the following months. A consumable launched in March that shows a rising repeat rate by June is building a base. One that sells hard on launch and flatlines is renting demand from ads.
Promo cohorts versus organic cohorts. Split the months you ran a deep Big Billion or Great Indian Festival push from the quiet months. Buyers acquired during a heavy discount window almost always repeat worse than buyers acquired at full price. The platform will not tell you this per customer, but the repeat-rate trend across promo-heavy and promo-light periods will.
Subscribe cohorts. If you sell anything repeatable, subscribe-and-save is the closest thing to a real customer cohort the marketplace will ever give you. Enrolments by month, and how many are still active three and six months later, is a near-clean retention curve. Guard it like the asset it is.

Each of these is a cohort in the way that matters. It groups demand by when and how it was acquired, then measures whether it persisted. That is the entire point of cohort analysis. Identity is a convenience, not a requirement.

The new-to-brand number is your acquisition denominator

Amazon’s new-to-brand metric is underused as a cohort input. Read alongside total orders, it tells you what share of this period’s sales came from buyers the brand had probably never seen. A high new-to-brand share with flat total sales means you are acquiring and leaking in equal measure. A falling new-to-brand share with rising sales means existing demand is carrying you, which is healthy until it is stagnant. Tracked monthly, new-to-brand becomes the front edge of every acquisition cohort you build.

Repeat-purchase rate is the signal to defend

If you only instrument one thing, instrument repeat-purchase rate at the SKU level and watch its trend. It is the cleanest retention proxy a marketplace gives most brands, and it answers the question that actually pays. Are we building something, or are we buying the same sale again every month.

The danger is reading it as a snapshot. A 22 percent repeat rate means nothing in isolation. The same number rising across three launch cohorts means your product and your post-purchase experience are earning a second order. The same number falling while ad spend climbs is the warning that you are acquiring worse buyers, or that a competitor undercut the reorder. This is the same trap we flag when teams stare at a flattering blended figure instead of the trend, and it is why a per-SKU read matters more than an account average. We have argued the same logic from the cost side in looking at profitability one SKU at a time, and retention and margin are the two halves of whether a SKU deserves its shelf.

Repeat rate also reframes acquisition. A SKU with a strong, rising repeat signal can justify a worse acquisition cost, because the second and third orders pay it back. A SKU that never repeats has to win on the first order or not at all. Without cohorts you cannot tell these two apart, and you end up funding both as if they were the same business.

From cohorts to the number that matters

A repeat curve is not the destination. It is the input to value. Once you have a defensible repeat-purchase trend and an average order value, you can build a rough estimate of what a marketplace buyer is worth over time, even though you can never see that buyer again. The estimate will be a range, not a point, and that is correct. It is still enough to decide whether a category is worth scaling.

This is where cohorts feed directly into estimating customer LTV on marketplaces. The repeat rate gives you the probability of a next order, the order value gives you its size, and the cohort trend tells you whether that probability is improving or decaying. Stack those and you have a lifetime-value band built entirely from signals the platform did not mean to give you. It will be coarser than a D2C model. It will also be the difference between guessing and reasoning.

Cohorts also tell you where to spend retargeting effort

The cohorts that repeat well are the audiences worth chasing back. Even without customer identity, the platforms let you reach lookalikes and prior viewers, and knowing which cohort actually returns tells you which intent is worth paying to re-engage. That feeds straight into retargeting marketplace shoppers when you do not own their data, where the whole game is spending re-engagement budget on the demand most likely to convert again rather than spraying it evenly.

What changed recently

The case for reading retention sideways got stronger over the last year, because the platforms themselves stopped pretending a subscription badge equals loyalty. In February 2026 Zepto quietly shut down Zepto Daily, its loyalty and subscription programme, ahead of its IPO, and Swiggy Instamart’s own chief described the market as so irrationally competitive that customers switch platforms without any real loyalty, per Inc42. The lesson for a brand is direct. A subscribe enrolment is still your cleanest cohort, but enrolment is not retention. Track how many of each month’s subscribers are still active at three and six months, because the platform will happily sign people up and just as happily let them lapse.

The second shift is where the platforms are actually investing their reporting effort, which is the ad side. Quick commerce ad spend on Blinkit, Zepto and Instamart jumped to roughly 4,000 Cr in 2025 and is projected near 6,000 Cr in 2026, and the platforms now hand brands granular shopper signals like basket mix, order timing, locality and purchase frequency to feed that machine, as Inc42 documents. Purchase-frequency reporting is a cohort input the moment you stop reading it as a vanity number and start grouping it by acquisition window. The data exists. Whether it changes a decision is on you.

The third shift is cost. Blinkit and Zepto hiked commissions through 2025 to push toward profitability, which means the bar for a SKU to deserve its slot just went up, per Business Standard. Higher take rates make the repeat curve the deciding variable, not a nice-to-have. A SKU that only ever wins the first order cannot absorb a richer commission. One with a real, rising repeat signal can. The math we used to treat as analysis is now the difference between a profitable line and a subsidised one, which is exactly the discipline behind quick-commerce unit economics after platform fees.

Make it survive contact with leadership

The honest weakness of marketplace cohorts is that they are proxies, and proxies invite an easy dismissal. Someone in the room will say this is not real cohort data, and they will be technically correct and practically unhelpful. Pre-empt it. State the proxy plainly, show the trend over enough months that noise washes out, and tie every cohort to one decision it changed. A cohort view that does not move a budget or a launch is decoration.

Three habits keep these cohorts credible.

Always show the trend, never the single number. One repeat rate is an opinion. Six months of the same cohort metric is evidence.
Name the proxy out loud. Say repeat-purchase rate as a stand-in for retention, not retention. The honesty is what makes leadership trust the rest.
Cut to the SKU and the channel. Blended cohorts hide the variance that is the entire reason to look. A healthy account average can sit on a hero SKU that repeats and a long tail that never does.

Getting this in front of decision-makers without burying them in tabs is its own craft, and it is exactly what we mean by a dashboard leadership will actually read. The cohort table is useless if it lives in a spreadsheet nobody opens.

The short version

Not owning the customer does not mean you cannot see retention. It means you have to read it sideways, from new-to-brand share, subscribe curves, and repeat-purchase trends grouped by how and when demand was acquired. Those proxies will never be as clean as a D2C cohort grid. They are clean enough to tell you which products earn a second order, which promos buy disposable buyers, and which categories deserve more money.

Our Analytics & Reporting work exists to assemble exactly these signals into cohorts a brand can act on, and our Marketplace Performance teams use them to decide where acquisition spend actually compounds. The customer is hidden from you. The curve is not. Build on the curve.