Overview

A DTC telehealth startup was spending across seven marketing channels with no clear picture of which were actually driving revenue and which were just taking credit for conversions that would have happened anyways.

The result:

• 3 findings that fundamentally change how the budget should be allocated and a reallocation recommendation estimated to generate +$42K/quarter from a single change.

View Attribution Dashboard → View Deep Analysis Dashboard

Dataset & Tools

Key Metrics

Three Findings That Change Budget Decisions

Attribution Model Comparison

All four models run on the same 467 conversions and $68,200 in revenue.

Paid Channel ROAS by Model

Funnel Analysis

Conversion Funnel — Drop-Off by Stage

Checkout Completion Rate by Channel

SEO Opportunity Sizing

Assumption: moving to top-3 position captures ~30% CTR. Organic CVR: 5.2%. AOV: $164.

Top two keywords alone = $324K/year in incremental organic revenue, both at LOW keyword difficulty.

Cohort Analysis

April's CVR is slightly lower, but the cohort converts 33 days faster and spends $12 more per order. Something changed between March and April — in targeting, onboarding, creative, or the product itself. That change is worth identifying and making permanent.

Demographic Performance

CVR & AOV by Age Group

The 45–54 group has the highest AOV ($173) despite a middle-of-pack CVR. They're the most valuable customers by revenue per conversion.

CVR by Device & Country

Australia's 14.2% CVR is notably below the rest. Worth investigating for UX or localisation issues. UK and Canada overindex on CVR relative to their revenue share — they're underpenetrated markets worth targeting.

Recommendations

What I'd Do With More Data

A few analytical caveats worth flagging:

• Uniform CPCs (~$3.16–3.38 across keywords) suggest broad match pooling in the ad click data, which limits keyword-level ROAS confidence. With exact-match data, the keyword analysis would be significantly more precise.
• The Markov Chain model assigns zero credit to Email and Paid Social, which is the technically correct answer given the journey data — but in practice, these channels likely have brand-building value that a 3-month window doesn't capture. A longer dataset would give a clearer picture.
• April cohort finding is promising but not yet causal. The 33-day improvement in conversion speed could reflect a targeting change, a product change, or simply seasonal effects. I'd want a 6-month cohort view before treating it as a reliable lever.
• Australia's low CVR (14.2%) warrants a session recording review — it could be a payment friction issue, a localisation gap, or just lower intent traffic. Hard to diagnose from aggregate metrics alone.

Methodology

<details> <summary><strong>1. Problem Definition & Requirements</strong></summary>

Outlined the core business questions: which channels drive conversions, how to allocate spend between awareness and retargeting, and how to properly credit multi-touch customer journeys. Translated marketing needs into data requirements covering sessions, conversions, events, ad clicks, and SEO keyword metrics.

</details> <details> <summary><strong>2. Data Simulation & Schema Design</strong></summary>

Prompt-engineered a synthetic DTC healthcare dataset using ChatGPT-5 to produce realistic GA4-style distributions across user demographics, session patterns, and conversion behaviours. Designed and validated a normalised Postgres schema (in Postico) covering: users, sessions, events, conversions, ad_clicks, semrush_keywords. Built three SQL helper views — session_channel_map, conversion_path, and campaign_costs — to power all downstream analysis without repeated joins.

</details> <details> <summary><strong>3. Exploratory Data Analysis</strong></summary>

Profiled all six tables in Jupyter Notebook. Key findings from EDA: user ages 18–69 (mean 43, US-dominant), Paid Social is the highest-spend channel, the 50% checkout abandonment rate immediately flags as a priority, and email-to-conversion paths are longer than other channels. This stage shaped the focus of the deeper analysis.

</details> <details> <summary><strong>4. Attribution Modelling — Baseline</strong></summary>

Implemented Last-Click, First-Click, and Linear attribution from scratch in Python. Compared credit allocation side-by-side to confirm that single-model reliance actively misleads budget decisions. Organic and Paid Social inflate under First-Click; Paid Search dominates Last-Click; Email emerges as an assist channel under Linear. No single model tells the full story.

</details> <details> <summary><strong>5. Attribution Modelling — Markov Chain</strong></summary>

Built a removal-effect Markov Chain model using Monte Carlo simulation (20,000 paths per channel). Process: computed transition probabilities from full session histories, modelled both converting and non-converting paths, then calculated each channel's incremental contribution by measuring the drop in overall conversion probability when that channel is removed from the model.

This is the most technically rigorous of the four models — it's the only one that actually measures incremental value rather than assigning credit by position.

Result: Paid Social drops to 0%, Display Retargeting rises to 24.3%, and Paid Search's true contribution is revealed to be significantly higher than Last-Touch suggests.

</details> <details> <summary><strong>6. Deep Analysis</strong></summary>

Extended across five dimensions after attribution:

• Funnel analysis — drop-off rates by channel and device at each stage
• Cohort analysis — conversion rates, speed-to-convert, and AOV by acquisition month
• Keyword ROAS — revenue per click across 2,019 attributed ad clicks
• SEO opportunity sizing — incremental revenue projections using CTR curves by position
• Demographic analysis — CVR and AOV segmented by age, device, gender, and country

</details> <details> <summary><strong>7. Outputs</strong></summary>

Delivered findings across four formats:

• Interactive HTML attribution dashboard (model toggle across all four models)
• Interactive HTML deep analysis dashboard (funnel, cohort, SEO, demographics)
• 9-page PDF report (consulting-style, with narrative, tables, and recommendations)
• 11-slide PowerPoint deck

</details>

Project completed September 2025. Dataset is synthetic and created for portfolio purposes.