A Shanghai consumer-goods company doubles email open rate with PIPL-compliant personalisation

Context

Your Shanghai consumer-goods company sells across 41 SKUs in personal care and home care, distributing through 280,000 retail points and three direct-to-consumer channels (Tmall, JD, WeChat Mini-Programs). Email and push-notification campaigns ran weekly with one creative variant per channel. Open rates on email had drifted from 18% to 11% over two years, push opt-out had risen to 6.4% per quarter, and the marketing team suspected the audience was being trained to ignore them. The CMO had a budget of CNY 12M (US$1.7M) per year on outbound marketing and could not show clear incremental return on most of it.

Challenge

Three constraints. First, customer data was fragmented across the three DTC platforms with no unified ID layer. Second, China consumer privacy regulations (PIPL) require explicit consent and clear data-handling disclosure for personalisation. Third, the in-house creative team produced content for the next campaign, with no time to produce six creative variants for personalisation experiments.

Approach

We ran a 4-phase model: identity, segment, personalise, hand-over. Identity (4 weeks) built a unified customer profile by reconciling identifiers across the three DTC platforms with PIPL-compliant consent capture. Customers without explicit personalisation consent were placed in a separate cohort and received the legacy non-personalised campaigns.

Segment (3 weeks) clustered the consented base of 4.8M customers into 9 behavioural cohorts using purchase recency, frequency, category mix, and channel preference. The cohorts were named in plain Mandarin for the marketing team (no model artifacts in cohort names), so anyone could explain the segmentation in a one-line sentence.

Personalise (10 weeks) deployed an LLM-assisted creative generator that produced 6 subject-line variants and 3 body-copy variants per campaign, with the in-house creative team approving every variant before send. Audience-cohort assignment to variant was algorithmic; final approval was human. The PIPL audit log captured every targeting decision.

Hand-over (parallel from week 12) trained two of your marketing operations team and one creative lead on the variant-approval workflow, the model-eval rota, and the cohort-refresh cycle.

Results

Email open rate rose from 11% to 21.3% across the personalised cohorts, with no improvement in the consent-withheld cohort (the control). Click-through rate rose from 1.7% to 4.2%. Push-notification opt-out fell from 6.4% to 2.1% per quarter. Attributed incremental revenue from the personalised campaigns came to CNY 38M (US$5.4M) over the first three quarters, a 3.2x return on the personalisation investment.

The PIPL audit completed without findings. The consent-withheld cohort gave the marketing team a clean control group for ongoing measurement, which the CFO cited as the deciding factor in renewing the budget for year two.

Lessons

The PIPL-compliant consent split produced a control group that turned every subsequent test into measurable evidence rather than vendor opinion. Naming cohorts in plain Mandarin removed the friction that usually kills marketing adoption of model-driven segmentation. Keeping creative approval in human hands preserved brand voice while still capturing the personalisation upside.