Data-Driven Attribution for Ecommerce in Pakistan

The Client

A Karachi-based direct-to-consumer brand selling educational toys, STEM kits, and family board games for children aged three to thirteen, founded in 2022 by two former schoolteachers who turned a weekend market stall into a recognized name among Karachi and Lahore parents. By the time they engaged WeProms Digital, the business was shipping roughly 1,740 orders a month at an average order value near PKR 4,000 — about 1,150 through the Shopify storefront, 380 through a Daraz flagship store, and 210 arriving as WhatsApp and Instagram messages that staff keyed in manually between other work.

The channel mix had grown the way most Pakistani D2C mixes grow: Meta carried prospecting and a thick retargeting layer, Google ran Performance Max and Shopping over the catalogue, Daraz sponsored listings topped up marketplace visibility, and a Klaviyo account sent campaigns to whoever landed on the list. Total paid spend sat around PKR 2.4 million a month. Cash on delivery accounted for 84% of orders, and 22% of first-time COD orders eventually returned uncollected — a fact the revenue reports never mentioned.

Growth had flattened after two strong Eid seasons. The founder’s question was simple and unanswerable with the reports she had: of everything we spend, what actually creates a customer we did not already have? Every monthly review ended the same way, because the platforms each told a different story and GA4’s default view sided with whichever touchpoint sat closest to the sale.

This engagement is an illustrative composite — a representative profile assembled from patterns we see across Pakistani D2C brands, with figures kept inside realistic ranges so your team can sanity-check fit against its own baseline.

The Problem

Four measurement failures were quietly steering the budget the wrong way.

• Last-click worship. GA4’s default non-paid last-click model credited 63% of conversions to branded search and direct traffic — the end of the customer journey, not the beginning. The response this produced was rational and destructive: pour budget into the channels that “converted,” trim the ones that did not.
• A quarter of orders no report could see. WhatsApp orders, Instagram DMs, COD confirmation callbacks that closed the sale, and Daraz sessions carried no source. At 28% of volume, the dark segment was large enough to invert every conclusion the dashboards supported.
• Platform claims that disagreed. Meta’s pixel, Google’s conversion counts, and GA4 differed by 18–31% on the same campaigns — partly consent and ad-blocker loss on Pakistani mobile traffic, partly a purchase event firing twice because the web tag and the Shopify webhook both claimed the same order, inflating platform counts by roughly 16%.
• Email reported as free money. Klaviyo showed PKR 1.1 million a month in attributed revenue against near-zero cost, when most of it was re-selling to audiences the paid channels had paid to acquire. The same revenue was effectively being counted twice and the cheaper claim was winning the budget argument.

The net effect: Meta prospecting had been cut twice in six months for posting “1.4x ROAS,” while PMax — riding branded queries and the catalogue overlap — reported 4.2x and absorbed every increase. Blended ROAS slid from 3.4x to 2.9x across two quarters of rising spend, and nobody could prove why with the numbers they had.

Phase 1 — Order-source audit and data foundation (Weeks 1-2)

No reallocation survives contact with a model nobody trusts, so the first fortnight went into one deliverable: a single order table in which every order — web, marketplace, or chat — carried a source, a customer status, and a delivery outcome.

One canonical order table. Shopify’s order record became the spine. Daraz settlement exports joined it weekly, keyed on product SKU bundles rather than platform order IDs, which disagree more often than they match. WhatsApp and Instagram orders were already keyed into the admin manually; the change was a mandatory source field on the entry form. The unified table fed a Looker Studio dashboard that became the only reporting surface in the business — a deliberate single-source-of-truth play that ended the parallel-spreadsheet era in week one.

Deduplicating the purchase event. The double-firing purchase event was the first fix inside GA4: the webhook send was made authoritative and the browser tag demoted to a fallback, which instantly reconciled most of the 16% platform inflation. A written UTM governance sheet and a locked link-builder followed, so campaign, audience, and creator links stopped arriving with mangled parameters that GA4 silently scattered across a dozen fake sources.

Source capture where the customer actually is. Two capture points covered the dark channels. The confirmation-call script — staff already call every COD order — gained one question: “Where did you first hear about us?” with six prompted options, answered at a 68% rate. The web checkout gained an optional one-field equivalent. Neither is statistically clean; both are dramatically better than nothing, and within two weeks the unknown-source share of orders had fallen from 28% to 11%.

Revenue at delivery, not at order. Because 22% of first-time COD orders returned uncollected, the table recorded delivered, confirmed, and RTO status, and every downstream number was computed on delivered revenue. In a COD market, a submitted order is a promise; weighting for delivery is the difference between optimizing the business and optimizing a report.

Channel	Platform / last-click claim	Validated picture after Phase 1
Meta prospecting	1.4x ROAS	Carrying 31% of first orders — the largest acquisition channel
Meta retargeting	6.8x	Overwhelmingly repeat buyers and cart-recoveries
PMax	4.2x	Brand queries and catalogue overlap inflating the claim
Google Shopping (non-brand)	2.6x	Genuinely productive commercial demand
Daraz sponsored	3.1x attributed	Incrementality untested, self-graded
Klaviyo	”Free” revenue	Assists on audiences paid channels acquired

The first-order column did what the dashboards never had: it stopped the argument about whether prospecting worked and started the conversation about how much budget it deserved.

Phase 2 — Attribution model build and validation (Weeks 3-5)

With clean order data underneath, the marketing attribution modeling work could begin — and the point of it was not a perfect number, but a defensible weighting that predicted what happens when money moves.

Model structure. GA4’s data-driven attribution served as the base model on a 30-day acquisition window, cross-checked against a custom position-based weighting run over the unified order table — necessary because GA4 sees none of the WhatsApp and Daraz journeys and under-credits view-assisted paths. Every conversion split new-versus-repeat, and PMax’s brand-query share was isolated through account-level brand-list controls so “PMax ROAS” stopped quietly including people who had already decided.

Validation by holdout, not by committee. A model earns budget authority by predicting reality. Two tests provided it. First, a 15% audience holdout on Meta prospecting ran for three weeks: new-customer orders in the holdout region ran 23% below the exposed group, meaning prospecting’s true contribution sat near twice its last-click credit. Second, Daraz sponsored listings ran a four-week alternating on/off cadence: incremental lift over organic baseline came in around 1.6x on sponsored spend — real, but far short of the 3.1x the marketplace dashboard claimed for itself.

The 2.1x threshold. The output every channel received was a validated marginal return estimate combining the model, the holdouts, and delivered-revenue weighting. Channels whose validated marginal return ran at least 2.1x the ones the budget would leave became the reallocation destinations. Meta prospecting and the non-brand Shopping layer cleared the bar comfortably; PMax ex-brand and the display layer did not.

By the end of week five the team had two numbers per channel it actually trusted — attributed first-order contribution and validated marginal return — and a rule for acting on them.

Phase 3 — Spend reallocation and creative absorption (Weeks 4-8)

The budget moved in three tranches of roughly 6–7% each, two weeks apart, so every shift could be read against delivered revenue before the next one went live. Moving the full 19% in one cut would have drowned the signal in delivery lag and made the lift indistinguishable from season.

Absorption before reallocation. The failure mode of every spend shift is starving or saturating the destination. Before the first tranche moved, Meta prospecting received three new creative concepts — unboxing films shot on phones, a “screen-free hour” angle for parents, and age-band targeting built from first-order data — and the Klaviyo account received the flows it had never had: welcome, post-purchase cross-sell keyed to the child’s age band, and a 60-day win-back. Budget followed capacity, not the reverse.

Spend bucket	Before	After	Why
Meta prospecting	19%	31%	Validated at ~2x its last-click credit by holdout
Meta retargeting	26%	21%	Mostly repeat buyers; capped at its real job
Google PMax + Shopping	38%	27%	Brand-query share isolated; display expansion off
Daraz sponsored	11%	11%	Held flat — modest incrementality, capped
Lifecycle email + creator seeding	0%	7%	Retention layer funded for the first time
Display and other	6%	3%	Failed both the model and the holdout bar

The reallocation totalled roughly PKR 460,000 a month — 19% of spend — out of retargeting, brand-inflated PMax, and display, into prospecting and a funded retention layer. Retargeting was not killed; it was capped at the job it demonstrably does, which is what made the politics of the move survivable inside the team.

Killing the free-money illusion. Klaviyo’s PKR 1.1 million was re-presented as assisted revenue tied back to the acquiring channel, not as a standalone free line. Nothing email earned changed; what changed was what the team credited — and once paid acquisition stopped being graded against a channel with no acquisition cost, the case for funding prospecting stopped needing an argument.

Phase 4 — Measurement cadence and compounding (Weeks 8-12)

Reallocation drifts back unless a routine holds it in place. The final phase installed the routine.

The Monday MER review. The dashboard’s first block became blended ROAS on delivered revenue, new-customer share, and cost per delivered first order — three numbers, one screen, reviewed weekly instead of monthly. Channel-level validated contribution refreshed monthly, when the holdout rotation turned over.

Biweekly reallocation triggers. Simple, pre-agreed rules: any channel whose validated marginal return stayed above 2.5x for two consecutive cycles became eligible for incremental budget; anything below 1.5x for two cycles took an automatic 25% cut pending review. The rules kept budget migrating toward the best channels without a re-modelling exercise every month, and they removed the personality from the decision.

A rotating holdout budget. One channel per month ran a small holdout or on/off test — a standing 3–5% of spend reserved for buying truth. Over the quarter this caught two drift events: prospecting efficiency decaying as audiences matured (refreshed with new creative) and the Daraz incrementality estimate sliding as organic marketplace rankings improved.

Peak planning. The engagement’s final weeks overlapped the run-up to the summer-holiday gifting peak, so the reallocation rules were stress-tested against a 2x spend scenario before the season hit: which buckets flex, which caps hold, and what evidence suspends a rule mid-peak. Deciding this in advance is what kept the peak from becoming an exercise in mid-season thrash — the pattern that had burned the previous year’s Eid budget.

Final Results

Measured at day 90 against the baseline month, on total spend held flat at PKR 2.4 million:

Metric	Baseline	Day 90
Blended ROAS (MER, delivered revenue)	2.9x	3.8x (+31%)
Monthly revenue	PKR 7.0M	PKR 9.1M (+30%)
Monthly order volume	~1,740	~2,260
Unknown-source orders	28%	9%
Cost per delivered first order	PKR 2,240	PKR 1,750 (-22%)
Email and retention revenue share	8%	15%
Double-counted conversion events	~16%	~2%

The trace: the revenue lift came from prospecting and retention receiving the 19% the bottom-funnel layers had been sitting on; the MER gain came from the same money producing more delivered first orders; and the unknown-source collapse came from two capture points — a confirmation-call question and a checkout field — that cost nothing but discipline. Every number in the table traces to a mechanism described in the phases, which is the standard any reporting should be held to.

What Made This Work

1. The order table came before the model. Modeling on top of unreconciled data produces confident wrong answers. Unifying web, marketplace, and chat orders — with delivery outcomes — is what made every later number worth arguing about.
2. Holdouts bought the budget’s trust. The team would have moved 5% of spend on a model’s say-so. It moved 19% because a three-week holdout showed new-customer orders falling 23% without prospecting. Evidence moves money that opinions cannot.
3. Absorption was planned, not assumed. New creative, new flows, and a funded retention layer existed before the budget arrived. Shifting spend into channels without capacity just buys more impressions of the same tired assets.
4. COD reality entered the math. Weighting for delivered revenue and RTO punished channels optimized for order placement rather than order completion — a correction that matters more in Pakistan than anywhere the checkout completes the payment.
5. Rules outlived the engagement. The trigger thresholds and the rotating holdout budget are mechanical enough to run without us, which is the difference between a reallocation and a quarter that happened to go well.

What Teams Can Apply

1. Count every ordering surface or you are reading a partial map. If WhatsApp, Instagram DMs, marketplace orders, or phone-back closures are part of your flow, a meaningful share of your revenue is invisible to GA4 by default. One mandatory source field — on manual entry, on the confirmation call, at checkout — recovers most of it for near-zero cost.
2. Reconcile before you compare. If your pixel, ad accounts, and analytics disagree by double digits, no attribution model fixes that; the disagreement just gets distributed more convincingly. Deduplicate the purchase event and crown one canonical order count first.
3. Split new customers from repeat before reading any ROAS. A channel can look efficient by reselling to people who were coming back anyway. First-order contribution is the number that tells you whether a channel builds the business or just services it.
4. Buy one piece of truth per month. A standing holdout or on/off test budget of 3–5% of spend converts attribution from a debate into a measurement. Run it against your most surprising claim first — that is where the money is usually misallocated.
5. Move money in tranches against delivered revenue. Three smaller shifts, each read against confirmed outcomes, beat one dramatic cut that cannot be separated from seasonality or delivery lag.
6. Grade owned channels on acquisition cost, not revenue. Until email and retention carry the cost of the audiences they monetize, every paid-channel comparison is rigged. This framework is built for the Pakistani ecommerce context — COD flows, marketplace overlap, and conversational ordering included — and the sequence holds across categories: reconcile, capture sources, model, validate, then move money.