Multi-Touch Attribution Case Study in Pakistan

The Client

A Karachi-based SaaS startup selling workforce-management software — HR records, attendance, and payroll — to Pakistani SMEs with 20 to 500 employees. The product runs on a self-serve motion with a 14-day free trial and a light-touch sales assist: a buyer signs up, and if the workspace looks promising, a salesperson calls to help with data import and payroll configuration. The company was eighteen months post-seed, running a three-person growth team, and spending about PKR 1.9M a month across Google Search, LinkedIn, Meta lead ads, and YouTube.

The numbers everyone saw monthly looked fine. Trial volume held around 540 starts. The blended cost per trial — PKR 3,100 — sat comfortably inside the benchmark range the team had set. Platform dashboards showed campaigns hitting their goals. The problem lived one layer down, in the definition of the goal itself. A trial start, it turned out, was nearly worthless as a unit. What mattered was activation: a workspace with five or more employees onboarded and a first payroll run completed. By that definition, paid acquisition was producing 64 paying-track customers a month at an effective PKR 29,700 each — a number no dashboard in the company could produce, because no dashboard joined the ad platforms to the product.

The CEO’s framing when the engagement began: “We can defend the trial number and we cannot defend anything else.”

The Problem

The deeper the audit went, the more the measurement layer — not the media buying — looked like the actual defect:

• Cheap trials, expensive customers. Meta lead-ads forms delivered trials at roughly PKR 1,900 each, a fifth of the cost of a LinkedIn-sourced trial. But activation rates diverged brutally: 11% for Meta lead-ads trials against 38% for LinkedIn and 42% for brand search. Corrected for activation, the cheap channel was the expensive one. No platform report performs that division, because no platform sees activation.
• Self-reported conversions that double-counted. Meta claimed 210 of the month’s 540 trials. Google claimed 260. Both were technically honest — view-through and click-through windows overlap — and jointly impossible. Budget meetings were being run on numbers that could not both be true.
• The real sales motion was invisible. Pakistani SME buyers research quietly and negotiate personally. Most closed deals passed through at least one phone call and a WhatsApp thread with a salesperson, all logged in Zoho but never connected to ad data. Attribution that ignores those touches mis-assigns credit for every deal they carried.
• Brand search absorbing the credit. Last-click handed 47% of conversions to brand terms — demand the brand would have captured regardless of ad spend — which made the bottom of the funnel look like the engine and the top look like a cost.
• One conversion, one value. A trial abandoned on day one and a workspace running payroll on day three counted identically in every report the team had.

This is the recurring measurement failure we see in Pakistani tech-startup marketing: teams optimise aggressively toward a proxy metric because the proxy is the only thing their tools can see. The engagement brief was to make the real metric visible — and then to move money according to it.

Phase 1 — Measurement Audit and Conversion Repair (Weeks 1-2)

The audit found the tracking in worse condition than the team feared, which paradoxically was good news: it meant the confusion was fixable at the instrumentation layer.

Event repair. The GA4 trial-start event double-fired on page refresh, inflating trial counts by roughly 9%. The event was rebuilt with a server-side confirmation call so a trial counted once, at the moment the workspace was actually created. The same server-side pattern was applied to the SQL and won stages pushed from Zoho, killing a class of duplicate and ghost conversions.

Defining and instrumenting activation. The single most consequential decision of the engagement: activation was defined precisely — five or more employees onboarded plus a first payroll run within 21 days of signup — and instrumented as a product event, then mirrored into GA4 and the ad platforms. From week two onward, every report in the company could segment by what a trial became, not merely that it started.

Stage architecture. Four revenue stages were locked as the shared vocabulary: Trial Start → Activated → Sales-Qualified → Won. Every touch, in every system, would carry the stage outcome of the account it belonged to.

UTM governance. A one-page tagging standard with a link builder replaced the team’s improvised conventions, so campaign naming finally matched across GA4, the warehouse, and the spreadsheets. Boring work; without it, the model in Phase 3 would be built on mislabelled input.

Repair item	Before	After (Week 2)
Trial events	~9% duplicated	Server-confirmed, deduped
Activation as a tracked event	Did not exist	Live in product + GA4 + ads
Revenue stages in shared vocabulary	None	4 stages, all systems
Campaign tagging compliance	Ad hoc	Standard, enforced

Phase 2 — Touchstream Unification in the Warehouse (Weeks 3-5)

With events trustworthy, the engagement built the thing the company had never had: one chronological record of every touch a buying account received.

Nightly unification. Spend from four ad platforms, sessions from GA4, opportunity stages from Zoho, and the sales-activity log — including calls and WhatsApp conversations, timestamped and linked to accounts through the call tracking and offline conversion integration pattern — landed nightly into a BigQuery project, keyed on a first-party identifier stitched across devices and joined to email at trial signup. The join covered a twelve-month lookback, long enough to see the multi-session, multi-week journeys SME buyers actually take.

Why the offline touches mattered so much. The first month of unified data showed that 100% of closed deals had at least one sales conversation between first touch and signature, and for 6 in 10, that conversation sat closer to the conversion than any ad click. Any attribution model excluding those touches was structurally blind to how this market buys. Ingesting them did not merely improve accuracy — it changed what kind of model was even possible, because the model could now see assists that had previously evaporated.

Coverage measurement. Touchstream coverage — the share of converting accounts whose full journey the warehouse could reconstruct — moved from 44% to 91%. The remaining gap was mostly pre-lookup dark traffic and some marketplace referral noise, documented rather than hidden.

Data source	Before	After (Week 5)
Ad spend (4 platforms)	4 separate dashboards	One fact table, nightly
Trial and activation events	Under- and double-counted	Server-confirmed, once
Zoho CRM stages	Disconnected	Joined to journey
Sales calls and WhatsApp	Invisible	Timestamped touches
Journey reconstruction	44% of conversions	91% of conversions

Phase 3 — Revenue-Weighted Model and Validation (Weeks 4-8)

Phase 3 is where the engagement becomes an attribution project rather than a plumbing project — the core of our marketing attribution modeling service.

Three models, built side by side. The team compared three views rather than trusting one:

Model	How it assigns credit	What it showed here
Last non-direct click	Last paid or organic touch before conversion	Still over-credited brand search (31%)
Position-based (40/20/40)	First and last touches weighted, middle split	Surfaced YouTube and content assists
Data-driven, stage-weighted	Regression on touch influence, weighted by progression to activation and revenue	Adopted as primary

The adopted model weighted each touch by its measured association with progression through the four stages — a touch that appeared disproportionately in journeys reaching activation and won stages earned more credit than one common to abandoned trials. The position-based model was retained as a cross-check, and the two agreed on channel ordering, which mattered for the trust the team would need to move money.

Validation before money. Two checks gated any reallocation decision. First, the model’s channel ranking was compared against twelve months of cohort data: for every historical acquisition cohort, did the channels the model favored actually activate at higher rates? They did — LinkedIn-sourced trials had activated at 38% against Meta’s 11% for a year. Second, a geo-split holdout ran on LinkedIn: spend continued in most of the country while two comparable regions went dark for four weeks. Treated regions produced measurably more activations per rupee, confirming causation rather than correlation. The model proposed; the cohort data and the holdout confirmed.

The first honest read:

• Brand search fell from 47% of credited conversions (last-click) to 15%. It closed demand; it did not create it.
• LinkedIn carried 33% of modelled influence at 24% of spend — the most underfunded channel in the mix.
• Meta lead ads held 21% of influence at 34% of spend. On pipeline-per-rupee, the destination channels beat it by 2.4x.
• YouTube and content placements quietly assisted journeys that brand search later closed.

The arithmetic that reframed the budget: every rupee moved from Meta lead ads and generic search into LinkedIn and comparison-intent search bought 2.4 times the activated pipeline. Roughly 23% of monthly spend was sitting in the wrong place.

Phase 4 — Staged Reallocation and Compounding Cadence (Weeks 8-12)

A model earns its keep only when it changes a budget line. Phase 4 moved the money carefully.

The reallocation. Over five weeks, in steps capped at 20% of any channel’s budget per week, the mix shifted:

Channel	Spend before	Spend after	Modelled pipeline-per-rupee
Meta lead ads	34%	16%	Baseline (weakest)
Generic non-brand search	27%	22%	~0.7x of destination set
LinkedIn sponsored content	24%	37%	2.4x baseline
Comparison-intent search (new)	0%	8%	2.4x baseline
YouTube	9%	11%	Assists heavily
Brand search	6%	6%	Held constant

The new comparison-intent campaigns — “best payroll software Pakistan,” competitor and alternative terms — were the purest expression of the model’s finding: buyers at the moment of comparison converted to activation at the highest rate of any intent band the team had ever funded.

Feeding activation back to the platforms. With activation and won events now trusted, they were imported back into Google and LinkedIn as offline conversions. The platforms’ optimisers, previously rewarded for filling the top with cheap lead-form submissions, began bidding toward buyers who run payroll. Within four weeks the platforms were doing the discrimination the team had been attempting manually.

The weekly cadence. A standing reconciliation reviewed model output against Zoho pipeline every week: channel-level cost per activated trial, coverage, and any cohort anomalies — before any further spend change could be approved. The rules were written down and enforced: no channel’s budget moves more than 20% in a week; every move is logged with the model finding and the validation evidence behind it; any metric whose warehouse number and platform number disagree by more than 10% freezes spend changes until the discrepancy is explained. Trial volume was allowed to drift down (it ended the quarter 4% lower) because volume was never the goal; activations were. The discipline sounds bureaucratic for a three-person team, and it is precisely what kept a mid-reallocation anomaly — a LinkedIn cost spike in week ten traced to a broad audience expansion the platform auto-applied — from being read as model failure and reversing the whole shift.

Board reporting. The monthly board pack replaced platform dashboards with three numbers by channel: cost per activated trial, paid-sourced new MRR, and payback. The “which channel produces customers” question, previously unanswerable, became a one-glance tile.

Final Results at 90 Days

Metric	Before	After	Change
Cost per activated trial	PKR 29,700	PKR 22,600	-24%
Paid-sourced activations per month	64	84	+31%
Spend reallocated to 2.4x channels	—	23% of budget	Staged over 5 weeks
Journey touchpoint coverage	44%	91%	+47 points
Brand-search share of credit	47% (last-click)	15% (modelled)	Honest read
Trial starts per month	~540	~518	-4% (deliberate)
Paid-sourced new MRR	Baseline	+29%	Quarter over quarter

These are illustrative outcome ranges built from the patterns WeProms sees across Pakistani B2B SaaS measurement engagements — not audited third-party figures. They exist so a growth team can judge whether an attribution build should plausibly return this shape of result at their spend level and sales-cycle length.

What Made This Work

1. The optimization target was the revenue stage, not the form fill. Defining activation precisely — five employees, one payroll run — and instrumenting it end-to-end changed every decision downstream. Attribution against trial starts would have kept funding the cheapest liar in the mix; attribution against activation exposed it.

2. The buying motion’s real channels were modeled, not assumed. Pakistani SME buyers negotiate on calls and WhatsApp. Ingesting those conversations as timestamped touches recovered the journeys of six in ten closed deals and made the model’s credit assignment match how revenue actually forms in this market.

3. The model was validated before the money moved. Cohort agreement and a geo-split holdout converted the model’s channel ranking from an opinion into evidence. That proof is what let a three-person team move 23% of a PKR 1.9M monthly budget without a board fight — and what made the moves stick when early trial counts dipped.

4. Reallocation was staged with hard guardrails. Capping weekly moves at 20% of any channel’s budget kept the funnel stable while the mix corrected. The deliberate trial-volume dip of 4% was the plan working: fewer tourists, more payroll runs.

5. Activation feedback closed the loop with the platforms. Importing activation and won stages as offline conversions let Google and LinkedIn optimise toward revenue-quality signals the team had previously hoarded in Zoho. The platforms became compounding allies instead of sophisticated volume merchants.

What Teams Can Apply

For Pakistani SaaS and B2B teams whose dashboards disagree with their billing system:

1. Define activation before building attribution. If you cannot state in one sentence what a “good” trial or lead does in week one, no model will save you. Instrument the activation event first; it is the denominator that makes every channel comparison honest.
2. Never reconcile platform dashboards against each other. Each platform counts favourably and on different windows; combined, they will always exceed reality. Land all conversions in one warehouse and count once, server-confirmed.
3. Join the CRM and the call log to the journey. If your buyers talk to sales before converting, those conversations are touchpoints with as much claim on credit as any ad click. An attribution model that cannot see your sales motion cannot describe your funnel.
4. Price your channels per revenue outcome, not per lead. Compute cost per activated trial or cost per qualified opportunity by source. In this market the cheapest lead channel is routinely the most expensive customer channel — the 1,900-rupee trial that never pays against the 5,800-rupee trial that does.
5. Move budget in stages, with evidence attached. Cap weekly shifts, log every move with the model finding and validation behind it, and reconcile against pipeline weekly. Attribution is not a quarterly reorganisation; it is a weekly operating rhythm.

WeProms Digital has applied this framework across Pakistani vertical SaaS, fintech, logistics-tech, and B2B service teams. The stages, platforms, and model choice change with each company; the sequence — repair events, unify the touchstream, weight by revenue, validate, reallocate in stages — stays constant.