How a leading subscription box brand boosted revenue and reduced churn with customer context

At a glance

• The core challenge: A leading subscription box brand had outgrown manual curation — rigid rules, no way to test what was working, and a growing catalog that the merchandising team couldn't keep up with.
• The approach: Built a two-sided recommendation engine using Faraday, scoring both customer affinity and product performance simultaneously, with subscriber records enriched from the Faraday Identity Graph (1,400+ data points across 240M U.S. adults).
• The results: ~5% lift in revenue per customer, ~3% reduction in churn, and 5x+ monthly ROI — across four consecutive months of algorithmic recommendations outperforming manual rules.

—

In subscription commerce, the wrong product doesn't just go unappreciated — it cancels subscriptions. And when a customer pays for a curated experience month after month, relevance isn't just nice-to-have. It's the entire product.

The math on retention makes this concrete: it costs 5–6x more to acquire a new customer than to keep an existing one, and returning customers spend 67% more on average than first-timers. A subscriber who receives consecutive boxes that don't match their lifestyle doesn't just disengage — they churn. And unlike a bad email, a bad shipment is hard to ignore.

That's the challenge a leading subscription box brand set out to solve. They had outgrown their manual curation process and needed a way to introduce real customer context into their monthly product matching at scale.

The problem: manual rules and missed opportunities

Before working with Faraday, this brand was manually assigning monthly boxes using written rules. The logic was informed by surface-level customer traits and preferences, but it was fundamentally rigid and prone to human bias. As the catalog and subscriber base grew, the cracks showed fast: the merchandising team found themselves buried under conflicting rules, manually balancing inventory constraints against subscriber preferences with no reliable way to test whether any of it was actually working.

They were operating on instinct rather than context — and leaving significant lifetime value on the table.

Bridging the context gap with FIG

To move beyond rigid rules, the brand needed to stop treating subscribers as simple profiles and start understanding them as fully realized households. First-party transaction data tells you what a customer bought. It doesn't tell you their household makeup, financial trajectory, or broader lifestyle affinities — the signals that actually predict what they'll want next month.

To fill that gap, the brand enriched their subscriber records using the Faraday Identity Graph (FIG): a continuously maintained dataset covering 240M U.S. adults across 1,400+ verified consumer data points, spanning demographics, financial signals, property data, life events, and lifestyle indicators.

Crucially, FIG also provides historical data — not just a current-state snapshot, but a full timeline for each data point. The model doesn't just know a household's income today; it knows how their lifestyle and spending patterns have shifted over the last five years. That trajectory is what lets a recommendation engine find combinations of signals no human rule-maker would ever think to look for.

The methodology: two-sided matching

With the enriched data foundation in place, Faraday implemented a two-sided matching framework — a meaningful departure from traditional recommendation engines that rank products by overall popularity.

Instead of one score, the model evaluates two vectors simultaneously:

• Customer affinity: How likely a specific person is to engage with a product
• Product performance: How well that product has performed with similar audiences

By evaluating both sides, the system predicts the mutual affinity between a person and a product — not just "is this product popular?" but "is this product right for this person, given everything we know about them and everyone like them?"

Behind the scenes, this is powered by explainable propensity models. The practical payoff for a brand: you can see exactly which data points are driving each product match — household signals, life events, spending trajectory — rather than trusting a black box.

Tuning for business reality

A common failure point for recommendation engines is ignoring operational constraints. An algorithm might identify the perfect product for a large cohort, but the warehouse might not have the inventory to support it.

Faraday solves this by letting the merchandising team manage global program constraints — inventory availability, product diversity, subscription cadence — rather than writing individual rules. The result is recommendations that are both deeply personalized and operationally feasible.

The results

To validate the approach, the brand ran a clean A/B test comparing Faraday's recommendations against their manual rules. And the results were clear:

• ~5% lift in revenue per customer
• ~3% reduction in overall churn rate
• 5x+ monthly ROI generated with Faraday
• Four consecutive months of algorithmic recommendations outperforming manual rules

These weren't marginal gains — they were statistically significant improvements that validated the shift away from manual curation. And because the models improve as more data flows through them, these aren't ceiling numbers. They're a starting point.

For a deeper look at how to predict and prevent churn across your subscriber base, see our complete guide to churn prediction.

Ready to optimize your recommendations?

If you want to talk through how a two-sided recommender could work for your subscription business, talk to a Context Consultant. Or if you'd rather get started on your own, try it on buy.faraday.ai.