Marketing Attribution Case Study in Pakistan

The Client

A Lahore-based SaaS startup building an AI-powered customer-support helpdesk for Pakistani and regional ecommerce companies. The platform routes tickets, drafts agent replies, and surfaces deflection analytics. The company had closed a Series A round nine months before the engagement and was in aggressive customer-acquisition mode, with a team of four in growth marketing and a monthly paid budget of roughly PKR 2.6M spread across Google Search, LinkedIn, Meta, and YouTube.

The product was strong and the team was capable. The problem was measurement. They were spending real money across five surfaces and could not confidently answer the most basic growth question: which channels actually produce a demo? Everything they reported to the board was filtered through GA4’s default last-click model, and the numbers that model produced did not match what the sales team experienced on calls. The result was a series of confident-sounding decisions that were quietly making acquisition more expensive.

The Problem: Last-Click Lying to the Team

Four symptoms kept recurring:

• Rising CAC, flat demos. Blended customer-acquisition cost had climbed 31% over two consecutive quarters — from roughly PKR 14,000 to PKR 18,400 per demo — while monthly demo volume had stalled at around 140. The team was paying more for the same outcome and could not see why.
• Branded search eating the credit. GA4’s last-click model credited 64% of conversions to branded search. This is the classic attribution illusion: a buyer sees a YouTube ad, reads a LinkedIn post, visits the site twice, and finally converts on a branded search. Last-click hands the entire win to the brand term, which makes branded search look like the company’s best channel and makes every awareness investment look worthless.
• Mid-funnel channels getting cut. Because video and display “didn’t convert” in the last-click view, the team had reduced YouTube spend by 40% and paused display retargeting entirely. Pipeline softened within a month — exactly the pattern you expect when you turn off the channels doing the awareness work — but the team could not prove the connection, so the cuts stayed.
• Fragmented data. Ad spend lived in five platform dashboards. Sessions lived in GA4. Demos lived in HubSpot. Revenue and churn lived in the billing system. Nothing was joined, so nobody could compute a true channel-level CAC or model how a touch in one platform influenced a conversion in another.

This is the pattern we see most often in tech-startup marketing in Pakistan: capable teams, real budgets, and a measurement layer that systematically misreads which work is producing results.

The cost of the misread was not just wasted spend. It was compounding misdirection. Because last-click credited branded search, the team kept increasing brand-term bids — bidding on demand the company would have captured anyway — while the channels generating that demand in the first place lost budget. Each cycle made acquisition slightly more expensive and the dashboard slightly more confident that the wrong channels were working. Breaking the loop required not a better campaign, but a better measurement system the team could actually trust to redirect money.

Phase 1 — Tracking Audit and Data Foundation (Weeks 1–3)

Before any model could be trusted, the raw data had to be trustworthy. The first three weeks were unglamorous and essential.

Server-side tracking repair. The existing GA4 setup was missing roughly one in four demo conversions — ad blockers, consent bounce, and a flaky client-side event on the booking confirmation page were losing signals before they reached the platform. We implemented a server-side tag layer that captured the booking event reliably and joined it to the ad-click context via a first-party identifier. This is the same GA4 setup and custom-configuration foundation we build for every attribution engagement, and it is the prerequisite for any model worth running.

CRM and billing integration. We connected HubSpot and the billing system into the same event stream, so a demo could be followed from first ad touch through to sales-qualified status, closed-won, and realised revenue. Critically, the offline conversions (a sales-qualified lead created in the CRM, a deal marked closed-won by a rep) were passed back to the ad platforms for optimisation — something the team had never done.

Touchpoint capture. Every ad click, organic visit, direct load, email click, and referral was stamped with a persistent anonymous identifier and logged with a timestamp, channel, campaign, and landing page. By the end of Phase 1 the team had, for the first time, a complete chronological touchstream for every account that eventually booked a demo.

The anonymous identifier was the quiet enabler of everything that followed. Pakistani SaaS buyers do not click one ad and book; they research across sessions, devices, and weeks, often starting on a phone and finishing on a laptop. Without a first-party identifier stitched across those touches, the warehouse could only ever see fragments of a journey, and any model built on fragments would reproduce the same last-click distortion the team was trying to escape. Identity stitching is unglamorous infrastructure, but it is the difference between an attribution model that tells the truth and one that simply lies more elaborately.

Data gap	Before	After (Week 3)
Demo conversions captured	~76%	~97%
Offline CRM conversions fed to ad platforms	None	SQL + closed-won
Unified touchstream per account	Did not exist	Live, 12-month lookback
Channels with spend + conversion joined	1 of 5 (Google)	5 of 5

The honest finding from the audit was that the team had been making six-figure reallocation decisions on data that was missing a quarter of its conversions. Fixing capture alone did not change spend — but it made everything that followed defensible.

Phase 2 — Attribution Model Design and Warehouse Build (Weeks 3–7)

With clean data flowing, we built the model that would replace last-click as the team’s source of truth. This is the core of our marketing attribution modeling service.

Marketing data warehouse. All spend, touchpoint, demo, and revenue data was landed nightly into a small warehouse built on BigQuery, transformed into a clean fact table, and surfaced through a Looker Studio dashboard. The team could finally see, in one place, what each channel cost, what it touched, and what it produced — by cohort, by campaign, by week.

Multi-touch model design. We did not hand the team a single off-the-shelf model. We built three and compared them side by side so the team could see how each reshaped the picture:

Model	What it does	What it revealed here
Last non-direct	Credits the last paid/organic touch before conversion	Still over-credited search; marginal change
Position-based (U)	40% first, 40% last, 20% split across the middle	Surfaced YouTube and LinkedIn mid-funnel weight
Data-driven (regression)	Weights touches by their measured influence on conversion	Most defensible; confirmed the U-shape

The team adopted the data-driven regression model as primary, with the position-based model as a stable cross-check. The two agreed closely, which built confidence that the reallocation decisions to follow were grounded rather than model-dependent.

We deliberately resisted the temptation to ship a single black-box model. A regression weight is only as trustworthy as the team’s willingness to act on it, and Pakistani growth teams — reasonably — do not move six-figure budgets because a dashboard told them to. Showing the same conclusion emerge from a rules-based position model, a regression model, and (in Phase 3) an incrementality test gave the team three independent lines of evidence. When all three pointed the same way, reallocation stopped feeling like a gamble and started feeling like a decision.

The first honest read. Under the new model, the channel mix looked nothing like the last-click picture:

• Branded search dropped from 64% of credited conversions to 19%. It was the closer, not the creator.
• YouTube rose from 4% to 22% of credited influence — the channel the team had cut by 40%.
• LinkedIn rose from 11% to 26%, validating the team’s instinct that professional buyers were researching there.
• Meta held roughly steady, quietly doing introduction work for smaller ecommerce support teams.

The team had been starving exactly the channels doing the hardest, most valuable work. That single realisation justified the entire engagement.

Phase 3 — Spend Reallocation and Testing (Weeks 6–11)

A model is only useful if it changes behaviour. Phase 3 turned insight into budget moves, carefully.

Phased reallocation. Rather than swinging the budget overnight, we moved roughly 23% of total spend across four steps over five weeks, restoring YouTube funding, increasing LinkedIn, trimming low-influence branded-search bids (which were already capturing demand the brand would have won anyway), and pausing two Meta campaigns that the model showed contributed almost nothing.

Incrementality guardrail. Models can be wrong, and the team was rightly nervous about trusting a new view to move real money. We ran two incrementality tests alongside the reallocation — a geo-holdout on the restored YouTube spend and a conversion-lift study on LinkedIn — to confirm that the channels the model credited were genuinely causing conversions rather than just correlating with them. This is the discipline from our incrementality testing and geo-experiments practice, and it is what separates attribution-driven budgeting from gambling on a dashboard.

The geo-holdout was the cleanest signal. We held YouTube spend at zero in two comparable regions while running it at full strength in the rest of the country, then compared demo volume from the held-out regions against the treated ones over four weeks. The treated regions produced demonstrably more demos, which confirmed the model’s claim that YouTube was doing real work. A model says a channel matters; a holdout proves it. Having both in hand is what let the team restore the YouTube budget they had cut without flinching.

Channel	Spend share before	Spend share after	Modelled efficiency change
Google (brand)	38%	24%	Already-captured demand
LinkedIn	19%	27%	+2.4x vs last-click view
YouTube	11%	21%	+2.4x vs last-click view
Meta	22%	18%	Flat to slightly negative
Direct/partner	10%	10%	Unchanged

Phase 3 results (by Week 11):

• Blended CAC fell from PKR 18,400 to roughly PKR 13,800 within six weeks of reallocation
• Monthly qualified demo volume rose from ~141 to ~186 at flat total spend
• Two incrementality tests confirmed the reallocated channels were causal, not correlated

Phase 4 — Measurement, Governance, and Compounding (Weeks 10–16)

The final phase turned the model from a one-off project into a permanent decision-making system.

Weekly cadence and governance. We installed a weekly attribution review where the growth team checked model output against incrementality and pipeline reality before approving any spend change over a defined threshold. The rule was simple: the model proposes, incrementality confirms, the team decides. No channel gets more than a 25% week-over-week change without a test behind it.

Compounding loop. As the team reallocated toward higher-influence channels and fed offline conversions back to the platforms, the ad algorithms optimised better, CAC kept falling, and demo volume kept rising. By the end of the first full quarter after reallocation the compounding was unmistakable.

Board reporting. The dashboard replaced a deck full of last-click charts. For the first time, the board saw channel-level CAC, modelled marketing-sourced revenue, and payback period in one honest view.

Final Results at One Quarter

Metric	Before	After	Change
Blended CAC	PKR 18,400	PKR 13,200	-28%
Qualified demos per month	~141	~196	+39%
Spend reallocated to modelled-better channels	—	23% of budget	2.4x efficiency
Branded-search share of credit (modelled)	64% (last-click)	19% (data-driven)	Honest read
Marketing-sourced revenue (modelled)	Baseline	+34%	Re-priced
CAC payback period	11 months	7.5 months	-32%
Demo conversions captured for optimisation	~76%	~97%	+21 pts

These are illustrative outcome ranges, built from the patterns we see across Pakistani SaaS and tech-startup teams, not audited third-party figures. They exist to help a buyer judge whether an attribution engagement fits their situation.

What Made This Work

1. The data foundation came first. No model survives bad input. Spending three weeks on capture, identity stitching, and offline-conversion feedback was the unglamorous work that made every later decision defensible. Teams that jump straight to “give us a multi-touch dashboard” build castles on missing conversions.

2. The model changed behaviour because it was stress-tested. Showing the team three models side by side — and confirming the chosen one with incrementality tests — built the trust needed to move real budget. A model nobody trusts changes nothing.

3. The reallocation was phased, not swung. Moving 23% of spend in steps, with guardrails, let the team learn without blowing up a working funnel. Aggressive overnight reallocation is the most common way attribution projects damage the pipeline they were meant to fix.

4. Last-click was the real enemy. The company was not under-investing in growth. It was systematically starving the channels doing the hardest, most valuable work because a default model told it those channels did not convert. Correcting that one misread produced most of the result.

5. Offline conversions fed back to the platforms. Passing sales-qualified and closed-won events back to the ad networks let the algorithms optimise toward revenue, not just form fills. This is the lever most Pakistani SaaS teams leave on the table.

6. The dashboard replaced the deck. Once the board saw channel-level CAC, modelled marketing-sourced revenue, and payback in one honest live view, the monthly reporting debate disappeared. Decisions that used to take a slide deck and a tense meeting started taking a glance at a tile. Measurement work pays its highest dividend not in the numbers it produces but in the arguments it retires.

What Teams Can Apply

For Pakistani SaaS and tech-startup teams whose channel mix does not match their intuition:

1. Distrust last-click from the moment spend exceeds one channel. If branded search is credited with more than about a third of conversions, the model is almost certainly lying. Build a multi-touch view before the lie gets expensive.
2. Fix capture before you build models. Audit how many of your real conversions actually reach your analytics. If the number is under 90%, you are optimising on a sample that does not represent reality.
3. Join your CRM and billing to your ad data. Until demos, pipeline, and revenue are in the same stream as spend, you cannot compute a true CAC — and the ad platforms cannot optimise toward revenue.
4. Confirm every model with at least one incrementality test. Models describe correlation; only a holdout or lift study proves causation. Treat the model as a hypothesis and the test as the verdict.
5. Reallocate in steps with a weekly cadence. Move no more than 25% of any channel’s budget in a week without a test, review the model against pipeline each week, and let compounding do the work.

WeProms Digital has applied this framework across Pakistani SaaS, fintech, ecommerce, and B2B service teams. The platform mix, sales cycle, and model choice change with each company — the foundation-first, model-stress-tested, phased-reallocation sequence stays consistent.