Cutting CAC 31% with ML Segmentation

The CFO Said No. She Was Right.

Q2 planning. Marketing team had built a case for a 40% increase in paid acquisition budget. The slide deck was polished. The CAC benchmarks were industry-standard. The growth projections curved upward satisfyingly.

Our CFO, Sunita, looked at it and asked one question: "What percentage of the users we're currently acquiring are profitable at 90 days?" Nobody in the room knew the answer. That was the problem.

We were spending ₹2.4 crore per month on acquisition without knowing which 33% of that spend was generating 89% of the value. Budget approved with conditions: prove channel-level LTV within 60 days.

That condition was the forcing function that built what became our most impactful data initiative of the year.

The Starting Reality

We Were Spending Money Blind

When we finally built the attribution model, the channel-level LTV data was stunning in how unequal it was. Same blended CAC across all channels. Completely different value being delivered.

Acquisition Channel	Monthly Spend	CAC	90-Day LTV	LTV:CAC	Profitability
Google Search (branded)	₹28L	₹210	₹890	4.2×	✓ Strong
Google Search (non-branded)	₹41L	₹380	₹720	1.9×	✓ Marginal
Facebook/Instagram	₹62L	₹295	₹310	1.05×	✗ Break-even
YouTube Pre-roll	₹35L	₹440	₹280	0.64×	✗ Losing
Affiliate Network	₹74L	₹160	₹195	1.22×	✗ Marginal

YouTube and Affiliate together represented 45% of our acquisition spend and were generating value at 0.64× and 1.22× LTV:CAC respectively. We were actively subsidizing unprofitable growth.

The CFO's instinct was right. We hadn't needed more budget — we needed to understand where the budget we had was actually working.

Why the Channels Differed

The User Quality Gap Was Behavioral, Not Demographic

We expected the high-LTV channels to have demographically different users: higher income, different cities, older. They didn't. The difference was entirely behavioral — intent level at the moment of acquisition.

Channel	D1 Depth (avg features used)	D7 Return Rate	90-Day LTV	Key Behavioral Difference
Google Search (branded)	3.8	61%	₹890	Already knew the product; came to convert
Google Search (generic)	2.9	44%	₹720	Research-intent; comparing options
Facebook	1.6	28%	₹310	Passive discovery; low initial intent
YouTube	1.2	19%	₹280	Passive awareness; rarely converts to habit
Affiliate	1.9	31%	₹195	Deal-seeking; high early churn

The differentiating factor was intent at entry. High-LTV channels brought users with existing intent. Low-LTV channels were trying to manufacture intent through exposure — which rarely produces durable user behavior.

The Reframe

Stop Optimizing CAC. Start Optimizing LTV:CAC.

The marketing team's north star had been CAC minimization. That metric, optimized in isolation, actively incentivized acquiring cheap, low-quality users from affiliate and display channels.

A ₹160 CAC from an affiliate user with ₹195 LTV is a worse investment than a ₹380 CAC from a search user with ₹720 LTV. The math is obvious. The dashboards had been hiding it.

We replaced CAC as the primary metric with LTV:CAC ratio and 90-day payback period. Every channel had to hit a minimum LTV:CAC of 2.5× to keep receiving budget allocation. That single metric change restructured the entire acquisition strategy.

The Framework

Channel Quality Scoring

We built a system that assigned every acquisition channel a quality score updated monthly, driving budget allocation algorithmically rather than through quarterly planning cycles.

User Quality Signals

D1 feature depth, D7 return rate, D30 transaction rate — all by acquisition channel. Tracked as cohorts, not blended. Channel quality score = weighted average of these three signals.

Budget Allocation Algorithm

Budget allocated proportionally to channel LTV:CAC ratio, with floor protection for strategic channels (brand, SEO). Channels below 2.5× threshold placed on 30-day improvement plan before budget cuts.

Lookalike Seeding

High-LTV user behavioral profiles (not demographics) exported to Meta and Google for lookalike targeting. Updated monthly. Each cycle narrowed the target to higher-quality incoming cohorts.

Real-Time LTV Prediction

ML model scoring new users at 72 hours with predicted 90-day LTV. Enables faster channel quality assessment than waiting 90 days. Channel quality updated from predictions, not just actuals.

Technical Build

Attribution Pipeline

The technical challenge was connecting acquisition channel (ad platform data) to 90-day revenue (product database) through a userId that persisted across anonymous → registered → transacting user states. That join was the hard part.

Attribution Stitching

→Anonymous device ID linked to acquisition UTMs at landing page
→ID stitching: anonymous ID → registered user ID upon signup via deterministic match
→Revenue events joined to acquisition source through persistent userId chain

LTV Prediction Model

→XGBoost model: 72-hour behavioral features → 90-day LTV prediction
→Training data: 14 months of matched acquisition-to-revenue pairs
→Output: predicted LTV per user, attributed back to channel + campaign + creative

Channel Quality Dashboard

→Real-time: LTV:CAC by channel (using 72hr predicted LTV)
→Lagging: actual 90-day LTV by channel cohort (monthly refresh)
→Budget reallocation proposals auto-generated weekly based on LTV:CAC movement

Execution

How It Unfolded

Weeks 1–3 · Attribution Build

Stitching the Data

→ Data confirmed: YouTube and Affiliate significantly below LTV:CAC threshold.

✓Built anonymous-to-registered ID stitching pipeline
✓Matched 14 months of revenue data to acquisition source
✓First channel-level LTV report produced

→ Data confirmed: YouTube and Affiliate significantly below LTV:CAC threshold.

Weeks 4–6 · Reallocation

Budget Restructure

→ Immediate CAC improvement as channel mix shifted to higher-quality sources.

✓YouTube spend cut 80% (from ₹35L to ₹7L monthly)
✓Affiliate network renegotiated: CPA model replaced CPL
✓Released budget redirected to Google branded and non-branded

→ Immediate CAC improvement as channel mix shifted to higher-quality sources.

Weeks 7–10 · ML Layer

Predictive LTV Scoring

→ First month: LTV:CAC blended ratio improved from 1.8× to 2.9×

✓72-hour LTV prediction model built and validated
✓Channel quality dashboard updated weekly from predictions
✓Lookalike audiences rebuilt from High-LTV behavioral profiles

→ First month: LTV:CAC blended ratio improved from 1.8× to 2.9×

Month 3+ · Compounding

Ongoing Optimization

→ Blended LTV:CAC reached 3.8×. Acquisition volume maintained. Spend down 31%.

✓Monthly model refresh with actual LTV outcomes
✓Affiliate partners tiered: top 12% (highest LTV users) got 3× commission
✓Programmatic display eliminated entirely

→ Blended LTV:CAC reached 3.8×. Acquisition volume maintained. Spend down 31%.

Results

Quarter-on-Quarter Impact

−31%

Monthly Acquisition Spend

same volume maintained

3.8×

Blended LTV:CAC

from 1.8× baseline

+67%

Profitable User Share

of new cohorts

₹1.8Cr

Monthly Budget Released

redirected to SEO + product

Metric	Before	After (90 Days)
Monthly Acquisition Spend	₹2.4Cr	₹1.66Cr (−31%)
Blended LTV:CAC	1.8×	3.8×
% Users Above 2.5× LTV:CAC	33%	71%
90-Day Payback Period	68 days	41 days
Budget on Profitable Channels	55%	89%

Insights

The Harder Lessons

CAC minimization without LTV context is self-defeating

Optimizing CAC alone will reliably drive you toward cheap, low-intent acquisition channels. Those channels produce users with poor engagement and high churn — which then requires more spend to compensate. It's a feedback loop that looks like growth but isn't.

The attribution join is the technical bottleneck

Most teams can't do channel-level LTV analysis because they can't reliably join acquisition source to long-term revenue. The anonymous-to-registered ID stitch is unglamorous data work — but it's the foundation everything else depends on.

Channel quality varies more than channel cost

YouTube CAC was ₹440 vs Facebook's ₹295. Facebook looked like the better deal. But YouTube LTV:CAC was 0.64× vs Facebook's 1.05×. Both were bad, but different bad. Same cost, different quality — and quality is harder to see.

Lookalike audiences from LTV data outperform demographic targeting

We rebuilt all our paid audiences from behavioral signals of High-LTV users rather than demographic profiles. The first cohort acquired via behavioral lookalikes had 2.3× better LTV than the demographic-targeted cohort at the same spend.

Failures

What We Got Wrong

✕ Cut affiliate spend too aggressively, too fast

Why it broke

Reduced affiliate spend 60% in Month 1. Acquisition volume fell sharply and we had no immediate alternative at scale. Revenue targets missed for 6 weeks.

What fixed it

Tiered approach: cut bottom 60% of affiliates by LTV quality. Retained top performers, increased their commission. Net spend fell 28% but volume recovered within 3 weeks.

✕ Used the same CAC target across all geographies

Why it broke

Metro cities had higher LTVs but also higher CACs. Our single CAC ceiling was constraining metro acquisition disproportionately — where the economics actually worked best.

What fixed it

Geo-specific LTV:CAC targets. Metro cities: CAC ceiling raised 40%. Tier-2: ceiling reduced. Blended ratio improved further as metro mix increased.

What's Next

The Next Layer

Creative-Level LTV Attribution

We can attribute LTV to channel and campaign. Next: attribute to individual creative. Which ad copy and visual format predicts highest-LTV users? Creative briefing informed by downstream revenue, not just CTR.

Real-Time Budget Reallocation

Currently monthly reallocation based on LTV data. Next: weekly automated budget shifts between channels within predefined bounds. Respond to quality changes in real-time rather than monthly planning cycles.

Lifetime Value Bidding

Feed 90-day LTV predictions into Google and Meta bidding algorithms. Bid not for the click — bid for the predicted downstream value of the user that click is likely to produce.

We didn't need more budget. We needed to know which 33% of our existing budget was doing 89% of the work — and then do more of that.

Building a Real-Time Intent Scoring Engine for Lead Conversion

Machine Learning

Predicting User Drop-Off Before It Happens (Churn Prevention Model)

Machine Learning

How We Predicted High-LTV Users Using Behavioral ML (100K+ Sessions)

10 min read

Machine Learning

Building a Real-Time Intent Scoring Engine for Lead Conversion

9 min read

Machine Learning

Personalizing Insurance Offers Using ML-Driven User Clustering

8 min read

All Case Studies

Browse all 30 breakdowns

Work With Me

Let's build something together

Connect on LinkedIn

Follow for more insights

Where to go next

Consumer App Growth Consultant
The consulting work this sits inside, with the rest of the evidence behind it.
App store conversion rate benchmark, subscription apps
Published figures, traced to their sources — and where the sources disagree.
Install-to-purchase conversion benchmark, travel
Published figures, traced to their sources — and where the sources disagree.
Product Growth Score
Thirteen questions that locate which part of the funnel is actually constraining growth.

Using ML to Reduce CAC by 31% via Smart Audience Segmentation

Results at a glance

The CFO Said No. She Was Right.

We Were Spending Money Blind

The User Quality Gap Was Behavioral, Not Demographic

Stop Optimizing CAC. Start Optimizing LTV:CAC.

Channel Quality Scoring

Attribution Pipeline

How It Unfolded

Quarter-on-Quarter Impact

The Harder Lessons

CAC minimization without LTV context is self-defeating

The attribution join is the technical bottleneck

Channel quality varies more than channel cost

Lookalike audiences from LTV data outperform demographic targeting

What We Got Wrong

The Next Layer

You might also like

Where to go next

The CFO Said No. She Was Right.

We Were Spending Money Blind

The User Quality Gap Was Behavioral, Not Demographic

Stop Optimizing CAC. Start Optimizing LTV:CAC.

Channel Quality Scoring

Attribution Pipeline

How It Unfolded

Quarter-on-Quarter Impact

The Harder Lessons

CAC minimization without LTV context is self-defeating

The attribution join is the technical bottleneck

Channel quality varies more than channel cost

Lookalike audiences from LTV data outperform demographic targeting

What We Got Wrong

The Next Layer