Driving Data Science Direction Independently — Impact & Ownership

Sure. About a year ago, while doing routine monitoring of our ad platform's performance metrics, I noticed something curious that wasn't on anyone's radar.

Situation: I was looking at our weekly advertiser retention dashboard and noticed that our new advertiser 30-day churn rate had been creeping up — from 35% to 42% over three months. Nobody had flagged it because the overall revenue numbers were still growing, driven by existing large advertisers spending more. The churn was happening among small and medium advertisers, and it was being masked by the top-line numbers.

Task: This wasn't part of any project I was assigned to. But I recognized that a 7-percentage-point increase in new advertiser churn, compounding over time, would eventually become a serious growth problem. So I decided to investigate.

Action: I spent about a week doing exploratory analysis on my own time, outside of my sprint commitments.

First, I segmented churned advertisers by industry, ad spend, campaign type, and onboarding path. I found that the churn increase was concentrated in advertisers who used our self-serve campaign setup — particularly those who launched their first campaign and saw low impressions in the first 48 hours.

I dug deeper and found the root cause: a recent change to our ad ranking algorithm had inadvertently raised the effective minimum bid for certain long-tail keywords. New advertisers with small budgets were setting reasonable bids based on our suggested ranges, but their ads weren't getting served because the actual clearing prices had shifted. They'd see zero or near-zero impressions, conclude the platform didn't work, and leave.

I put together a concise analysis deck — three slides: the trend, the root cause, and three proposed solutions:

1. Update the suggested bid ranges to reflect current market prices
2. Add a "your bid may be too low" alert in the first 24 hours
3. Give new advertisers a small impression boost in their first 48 hours

I shared this with my manager first, then presented it to the Ads product leadership in their weekly review. The VP of Ads actually paused the meeting and asked the ranking team to prioritize investigating the bid threshold issue.

Result: The ranking team confirmed the minimum effective bid shift within a week. We implemented solutions 1 and 2 within a month. The 30-day new advertiser churn rate dropped from 42% back to 36% within the next quarter.

Back-of-envelope, recovering those advertisers was worth approximately $3M in annualized revenue from the SMB segment alone.

Two things. First, I believe that a data scientist's job isn't just to answer questions — it's to ask the right questions that nobody else is asking. Dashboards show you what you've decided to measure. The most important findings are often in the things you haven't decided to measure yet.

Second, I had the context to connect the dots. I was seeing the ranking algorithm changes in one set of meetings and the retention metrics in another. Nobody else was in both rooms. Being at that intersection is one of the unique advantages of a data scientist embedded in a product team, and I think it's my responsibility to use that perspective proactively.

I was honest with my manager. After the initial exploratory week, I said: "I found something that I think is costing us millions in advertiser retention. Can I formally allocate 30% of my time for the next two weeks to build a rigorous analysis?" She agreed immediately because I'd framed it in terms of business impact, not "I find this interesting."

I think the key is: don't ask for permission to explore, but DO ask for permission to go deep. The initial investigation was a few hours of curiosity. The deep-dive and presentation required real time commitment, and that needed alignment with my manager.