Home › Case Studies › Multi-Touch Attribution Analysis

Multi-Touch Attribution Analysis

Worked sample: the company and data are fictional. The model and method are the ones I use in my work.

A side-by-side comparison of four attribution models across $4.2M in closed-won revenue, plus a budget reallocation recommendation built for CFO sign-off. Built to settle the perennial Sales-vs-Marketing argument about credit.

The problem

Marketing was running first-touch and crediting paid social with 35% of revenue. Sales was running last-touch and crediting outbound with 60%. The CFO was sitting between them with a $1.8M marketing budget asking which channels to fund next quarter, and getting two different answers depending on who was at the table.

The approach

Pull a clean 12-month closed-won dataset with every touchpoint timestamped: paid, organic, outbound, events, content. Run all four standard models against the same data: first-touch, last-touch, linear, and W-shaped (with weight at first-touch, opportunity-creation, and closed-won). Show the contribution per channel under each model, side by side. Layer in cost data to compute channel ROI and CAC, not just attributed revenue.

What the model is built to do

• Gives the CFO, CMO, and CRO one chart they can agree on, even when each prefers a different model.
• Supports a budget reallocation recommendation: defund two underperforming paid channels and shift ~22% of the marketing budget into content and events, where the W-shaped model shows the strongest mid-funnel lift.
• Answers "are we overspending on attribution-credited channels with poor CAC" with channel-level evidence: in the sample data, two channels qualify for a pullback.
• Sets up a follow-on test: track pipeline-to-close conversion the quarter after the reallocation to confirm the lift the model predicts.

What I would do differently

Bring Marketing into the model selection early. Marketing teams have a legitimate preference for first-touch because it values their top-of-funnel work, and dismissing that preference as bias misses the point. The right answer is multi-model attribution as a standing report, not picking one model and forcing everyone to use it.