A 2x to 5x overcount corrected, and breach detection moved from weekly to same-day
A scaling direct-to-consumer meal brand
The Brief
The brand was scaling paid acquisition toward an aggressive goal: 1,000 new customers per week at or below an A$80 blended new-customer CAC, with spend running through several media buyers at the same time. Three problems sat underneath it.
First, the numbers could not be trusted. Platform-reported conversions overcounted new customers by 2 to 5 times against verified store orders. Broken UTM parameters dumped spend into unattributable buckets, so per-campaign CAC could not be measured at all. Attribution was split across three surfaces, each telling a different story, with no agreed source of truth.
Second, there was no common standard to judge performance. Each media buyer reported their own way, runaway campaigns were not caught until the weekly review, and there was no objective, reproducible line between an operator worth more budget and one worth cutting.
Third, the pieces did not connect. Acquisition cost, lifetime value and margin lived in separate places, so no one could see on a single reproducible view whether a dollar of spend was actually paying back.
The same week, counted two ways. Where the platform and the store conflicted, the overcount was discarded rather than blended into an average of two figures that were both wrong.
The Approach
One source of truth Verified store orders on last non-direct click became the only accepted count of new customers. No platform-reported conversion figure reaches any headline number.
Reconcile, do not average Where platform and store figures conflicted, we discarded the platform overcount rather than blending two figures that were both wrong. An average of a right number and a wrong number is a wrong number that looks defensible
One standard for everyone Every campaign and every operator is judged against the same CAC caps and the same single-variable-test discipline.
Catch it the same day Breach detection moved from the weekly review to a daily cockpit.
Follow the money end to end Every acquisition limit is anchored to a CLTV-derived maximum CAC, and acquisition cost, lifetime value and margin are tied into one return-on-investment view.
What we Did
Built a verified attribution layer. Restored UTM integrity and locked store last-non-direct-click as the single source of truth, so spend and new customers roll up cleanly by campaign and by operator. This foundation feeds every tracker below, which is why it had to be fixed before anything else was worth building.
Built a daily control cockpit. An intraday view reads each spend cohort at set times through the day, auto-computes CAC against its cap, flags any breach, and assigns the action — hold, trim, or cut — the same day. A runaway campaign is caught the day it happens, not at week’s end.
The cockpit reads each spend cohort at set times through the day, so the decision lands while the spend is still live. Up to six days of overspend removed from every breach.Built a CAC tracker with a bucket engine. A weekly panel that splits the account into proven scaling, owner tests and mechanical tests, so blended CAC is never read as one misleading number. Plus an operator scorecard that qualifies, graduates or removes each media buyer against the same standard.
Read straight from the CAC tracker. The blended line tracks close to the cap while the machine and the test buckets diverge by two orders of magnitude. A log scale is the only way all three fit on one axis
Qualification, graduation and removal all run off the same verified CAC figure. One standard for every media buyer, applied the same way every week.
Built a CLTV tracker. A nine-tab rolling tracker across 12-month and 24-month windows, with segment and product-line breakdowns and unit economics, refreshed weekly. Delta-correction means week-over-week movement reflects real change rather than noise from overlapping data exports.
The highest-leverage retention lever, isolated. 43.5% of new customers reach a second order, down from 47.9% a year ago. Nineteen days to a second order on average, eleven at the median.Built an ROI tracker. Ties acquisition cost, lifetime value and margin into one return view, so every operator and product line is judged on whether spend actually pays back, not on raw return on ad spend.
Operator ROI calculator. Inputs: AOV A$139, 38% gross margin, A$23 delivery and packaging, A$676 twelve-month CLTV across 4.95 orders. Judged on the first order the same customer looks like a loss.Diagnosed product lines and retention levers. Classified revenue and lifetime value by product line, mapping by SKU first with a keyword fallback, then ranked the retention levers that move lifetime value, so work targets the single highest-leverage gap rather than spreading thin.
Results in Detail
Verified store orders established as the single source of truth, correcting platform overcounting that had inflated new-customer counts by 2 to 5 times, with per-campaign CAC measurement restored by fixing UTM integrity.
Nothing downstream was worth building until this held. Every CAC read, every operator judgment and every payback calculation inherits this number.
Breach detection moved to same-day. A campaign running over its cap is caught and actioned the day it breaches, rather than surfacing at the weekly review.
On a weekly cycle a runaway campaign spends for up to six more days before anyone sees it. That gap is the whole cost of the old process, and it recurs on every breach.
Every media buyer on one hard standard. Operators are qualified on entry, graduated after three clean wins, or removed when they miss the CAC cap. Over the engagement, several operators were cut against this standard and budget was concentrated on the proven ones.
The proven scaling machine separated from deliberately-expensive tests, showing the underlying machine runs at a fraction of the blended CAC once tests are isolated — a read that a single blended number had hidden entirely.
Acquisition cost, lifetime value and margin connected in one ROI view, which showed a contribution payback of about 2.2 times at the affordable CAC ceiling.
Read on the first order alone, the same customer looks like a A$37 loss. The difference between those two readings is the difference between cutting a working channel and scaling it.
The single highest-leverage retention lever quantified. Moving a customer to a second order is worth about A$154 more per customer, a 116% lifetime-value uplift, which set the priority for all downstream retention work.
Why the blended number was the dangerous one
- Blended CAC sat near the cap and looked broadly healthy. Underneath it, the proven machine and the test buckets differed by two orders of magnitude.
- Judged on the blend, the machine looks unremarkable and the tests look like failure. Judged separately, one is worth scaling and the other is worth exactly what it costs to learn from. A single average would have led to cutting the wrong one.
Tools & technologies
What the build runs on
verified