Case Studies

Challenge

The client needed to know how much CRM communications (SMS, push, email) contribute on top of purchases customers would have made anyway

What we did

• Designed a global holdout group — customers who received no communications for a month
• Checked group balance on historical data before the test started
• Found a selection effect: when uploading to the messaging service, 52% of customers dropped out of the test group versus 18% of the control group, and those who dropped out were less active
• Recalculated the effect on the original random split and separately ran a segment‑level SMS A/B test

Outcome

The company now has a working estimate of CRM incremental revenue and a scaling criterion: the effect must cover discounts and channel costs. Instead of messaging everyone, the team now knows exactly whom to contact

Challenge

Report figures didn't match the accounting system, marketplace, wholesale and CRM data lived in different places, and weekly commercial reporting was compiled by hand

What we did

• Integrated marketplace data, wholesale data from the ERP, the order management system and the CRM platform into the warehouse
• Built data marts for revenue and cost of goods, year‑over‑year sales, digital goods and pre‑orders
• Launched a group‑wide glossary and agreed on disputed definitions (e.g., “new release”)
• Built an accounting data quality dashboard and started a data lineage map

Outcome

Revenue and cost of goods in the data marts can be trusted without manual reconciliation; weekly commercial reporting is built on the same data, and the question “where does this number come from?” has a documented answer

Challenge

Acquisition reports undervalued paid search, and recalculating attribution for returning users took half a day, slowing down budget decisions

What we did

• Audited link tagging and found visits with an untagged parameter that were not credited to any source
• Fixed the tagging and recalculated a year of historical data
• Rewrote the attribution logic for returning users
• Configured the web analytics tag on the sign‑in domain so sessions no longer break, and set up offline conversion uploads to ad platforms

Outcome

The channel had been performing better than the reports showed all along. Budget decisions now rely on accurate numbers, and attribution is recalculated in a short window instead of half a day

Challenge

The product team wanted to run a blanket 20% discount on in‑game items that activate immediately after purchase, expecting sales to grow

What we did

• Built a saturation model: on what share of play days active players already have their key slots filled with items
• Split revenue into direct purchases and purchases through bundles and tiers
• Compared churn among players who buy these items with the rest of the player base
• Modeled discount scenarios for each slot type

Outcome

The model showed that the discount would push most slots into the red: a saturated core won't play more, so revenue would simply drop by the size of the discount. The discount was not launched; instead, we recommended targeted discounts only on variable slots, where collecting matters more than the timer. Along the way, we showed that these items drive retention rather than short‑term revenue

Challenge

The team needed to understand whether to invest in mechanics that bring players together, and where players drop off before joining the core audience

What we did

• Built a social activity dashboard across several titles — over a million rows of data
• Compared players in and outside guilds by spend, purchases and payer share
• Ran a cohort analysis of churn before players reach the core audience
• Built a report on player gear and progression and automated data collection

Outcome

Social mechanics came to be seen as a lever for the game's economy rather than social features for their own sake. The cohort analysis identified a group of players with accelerated churn at a specific lifecycle stage — a clear target for product changes

Challenge

Requests came in through different channels, turnaround times were unpredictable, reports and research were scattered, and internal stakeholders didn't know what they could bring to the analysts

What we did

• Surveyed more than 50 internal stakeholders and drew up an improvement plan covering speed, data quality and transparency
• Moved all requests into a single tracker and defined request types and SLAs
• Built a wiki: a report catalog, a research catalog, a methodology standard and a glossary
• Assigned a lead and a backup analyst to every project, introduced a single SQL style guide and rules for working with AI
• Launched a monthly digest written in plain business language

Outcome

Analytics became a predictable service: it's clear where and how to submit a request, how long to wait for an answer and where to find existing reports. Outdated reports were archived and database load was reduced