How a DTC skincare brand achieved 18 percent order value growth and 2.4x email conversion through AI personalization
A seven-year-old DTC skincare brand converted personalization from marketing concept to operating system. Within 90 days, order value rose 18 percent, email conversion multiplied 2.4x, and repeat cycles shortened 18 percent. The driver was not a specific algorithm; it was consolidating scattered data and having engineers work closely with the marketing team. This breakdown explains what was built, where problems emerged, and what approaches were deliberately avoided.
By
Tenten AI 交付團隊
產業交付
Published
November 5, 2025
Read time
6 分鐘
The results were clear within 90 days of launching AI-driven personalization and segmented email campaigns. Order value rose from NT$1,180 to NT$1,392 (up 18%), email checkout conversion improved from 1.1% to 2.6% (a 2.4x increase), and repeat purchase cycles shortened from 74 days on average to 61 days.
This case examines how a skincare brand transformed personalization from a marketing term into an operating system. The marketing team needed to trust it and use it every day. That transformation moved the numbers.
The starting point: plenty of data, zero personalization
The brand had tools: Shopify, Klaviyo, GA4, and a purchased plugin that claimed to 'AI-select products.' The homepage displayed the same six bestsellers to every visitor. Email campaigns broadcast identical messages to 300,000 subscribers. Customer service had no visibility into the fact that someone on chat had returned an item the previous week.
Systems and data sources operated separately. Purchase history lived in Shopify, browsing behavior in GA4, returns in the support system, and skin questionnaires in an unused Google Sheet. 'Personalization' consisted of inserting first names into emails: 'Hi {first_name}'.
The team skipped building models in week one. Instead they consolidated scattered datasets into a single customer database, giving each person a queryable, complete profile. This groundwork was essential; without it, downstream work would fail.
What the team built
Customer segmentation came first. Rather than deploying deep learning immediately, they combined RFM (recency, frequency, monetary value) with skin-type data and ingredient preferences to create 14 business-meaningful segments. The segments were designed for marketers to understand and write for; for instance, 'combination skin, purchased serums in the past three months, sensitive to alcohol.' The marketing team needed to grasp these groups well enough to produce copy.
Personalized recommendations moved to real-time calculation. Collaborative filtering identified products that previous buyers wanted (people who bought A also wanted B), combined with content similarity based on ingredients and skin type. Business rules for inventory and margin prevented the system from pushing out-of-stock or razor-thin-margin items indefinitely. Every recommendation was built to be explainable; marketing could see why each product was suggested, which meant they trusted putting it on the homepage.
Email shifted from broadcast to triggered journeys. This yielded the largest conversion gains. Automated sequences replaced single campaigns: replenishment reminders based on individual usage patterns, cross-category suggestions, win-back emails after returns, and exclusive early access for high-value customers. The unified customer database generated email content and product recommendations dynamically.
Results showed in the metrics.
| Metric | Before launch | After 90 days | Change |
|---|---|---|---|
| Average order value | NT$1,180 | NT$1,392 | +18% |
| Email checkout conversion | 1.1% | 2.6% | ×2.4 |
| Repeat purchase cycle | 74 days | 61 days | −18% |
| Email unsubscribe rate | 0.9% | 0.5% | −44% |
| Recommendation click revenue share | Not applicable | 31% of revenue | New metric |
The unsubscribe rate fell by nearly half. Segmentation meant customers received fewer emails but more relevant ones. The healthier list grew. Personalization typically affects conversion rates; what it accomplishes first is protecting sender reputation and deliverability.
Where the team stumbled
Not everything went smoothly.
The replenishment reminder algorithm was initially too simple. Usage cycles were calculated from purchase intervals alone, which created constant 'your item is running out' messages for customers who stockpiled. Complaints arrived within two weeks. The team then incorporated actual product sizes and recommended usage rates, which improved predictions.
The team declined to build real-time personalization on the homepage. While technically feasible, the brand had only 200-plus SKUs; per-user customization would not justify the added complexity. Personalization depth should align with product catalog size and data volume. Over-segmentation of thin customer groups degraded rather than improved predictions.
The most time-consuming phase was building marketing team confidence. When the recommendation engine launched initially, they reviewed each suggestion daily, concerned about awkward or off-brand combinations. The engineering team embedded for three weeks and built a control panel where they could suppress individual recommendations or override rules. After that, the marketing team operated the system fully.
One-sentence summary
Numbers moved not because of a clever algorithm. They moved because the team unified scattered data into one coherent customer view and remained involved enough for marketing to operate safely. Building an impressive presentation on personalization is routine. The difference appears when it runs continuously for 90 days with daily use and budget allocated toward it.
The engineering team was stationed on-site during implementation. They monitored adoption metrics and learned alongside the marketing team instead of simply delivering API documentation. The payoff from personalization begins only when someone actually uses it.
One stuck workflow
is enough to begin
Tell us what the team does today, where it breaks down, and what a better working day should look like.