BeautyPlus AI Filter: 0-to-1 GTM Launch
Context
BeautyPlus had built its reputation on traditional beauty retouching — skin smoothing, facial reshaping, filters. In early 2023, generative AI was breaking into mainstream consumer apps fast. Competitor apps and creator accounts in Japan were already pushing AI-styled photo transformation and gaining traction on TikTok and Instagram.
The strategic question wasn't whether to launch AI filters. It was whether we could move fast enough to matter. The standard client-side development cycle at BeautyPlus ran on a quarterly release cadence — by the time a native feature shipped, the trend window would already have closed.
I was the PMM leading overseas GTM strategy of this project. My scope covered go-to-market planning, in-app channel management, localization for East-Asia and North America markets, and commercial growth across the full product lifecycle.
The Challenge
- The harder, ongoing challenge wasn't reading the market. Japan's TikTok ecosystem and competitor moves gave us clear signals on which aesthetic directions were gaining momentum. The real difficulty was what came after: continuously mapping those signals to our existing user base, judging which effects would land in which markets, and making that call quickly and repeatedly as new AI styles emerged week over week. Our native app release cycle made this even harder — shipping on a quarterly schedule meant we were always behind.
- Building a monetisation model that didn't cannibalise existing revenue. BeautyPlus's primary business was subscription-based retouching. We needed to find a way to generate incremental revenue from AI effects without eroding the subscription metrics that already drove the business.
Phase 1
Bypass the release cycle — launch as H5
The decision to build and ship as an H5 MiniApp rather than waiting for native client integration was one I pushed from the marketing side. An H5 web-based flow embedded in the app could be built and deployed in weeks, not months. It also meant full funnel instrumentation from day one — we could track every step from entry to save, and iterate on both the product flow and the promotion strategy in near real-time.
The H5 funnel had seven steps: enter → browse effects → upload photo → confirm generation → wait → save → share. We tracked conversion at each layer. This gave us the data layer we needed to make fast decisions.
Cold start promotion strategy: To validate the concept quickly, we ran deliberately interruptive in-app promotions. Pop-up windows were set to fire at 8pm local time on weekends in the US, timed to peak usage hours when DAU was highest. The approach was intentional — we weren't optimising for user experience yet, we were optimising for data.
Phase 2
Expand fast, promote smarter
Once the product flow was validated, Phase 2 shifted to velocity. The generative AI model landscape was moving quickly, and new aesthetic styles — pixel art, clay, retro manga — were gaining traction as fast as older ones plateaued. We partnered closely with the design team to ship new effects as quickly as the underlying models allowed, using novelty itself as a retention mechanism.
With more data on effect-level performance — including usage volume, save rate, and repeat engagement by market — we could adjust promotional weight by effect rather than treating all content equally. High-save-rate effects got banner and pop-up placement, while lower-performing effects were deprioritised or eventually delisted.
Japan-specific localisation became a deliberate commercial decision. Japanese market data pointed toward portrait-realism and stylised illustration aesthetics over the bold Western-facing styles that performed well in English-speaking regions. We built separate effect ordering configurations for Japan and ran localised promotional creative.
1M+
DAU within first month
5%+
Penetration rate
8+
Effects tried per session (up from ~6)
~2.5%
Generation abandonment (down from ~3%)
Phase 3
Introduce incremental monetisation without cannibalising subscriptions
The commercial challenge in Phase 3 was structural. BeautyPlus's subscription model was built around core retouching features. A large portion of the AI filter audience were free users engaging with the novelty of the effects, not necessarily looking for a long-term feature they'd pay a monthly fee to access.
The approach was to introduce a credits-based bundle system alongside subscriptions, allowing users to make one-time purchases for AI content without requiring a full subscription commitment. Subscriber benefits were structured deliberately — subscription users got access to a base tier of AI effects as part of their existing plan, creating a conversion pathway from free AI users into subscribers over time.
Channel economics mattered more than channel volume. We managed three in-app channels:
- Pop-up windows — ~22% click-through rate
- Homepage rotating banners — ~1% CTR
- MiniApp navigation tab — ~0.4% CTR
The Halloween seasonal campaign became the proof-of-concept for the paid bundle model at scale. We launched 13 Halloween-themed effects as a limited-time bundle, drove the campaign through pop-ups and banners timed to peak engagement windows, and made a deliberate call to keep video effects minimal — early data showed video conversion was dramatically lower than photo, primarily due to the friction of finding and uploading a sub-10-second clip.
80%
MoM subscription lift during Halloween campaign
22%
Halloween bundle share of AI MiniApp credit revenue
2% → 22%
Free-to-subscriber conversion in 6 weeks
+10%
Q4 homepage revenue vs prior quarter
Results
| Metric | Result |
|---|---|
| DAU at first-month peak | 1M+ |
| First-month penetration rate | 5%+ |
| YoY subscription growth | 2× (100%) |
| Holiday campaign MoM subscription lift | 80% |
| Free-to-subscriber conversion | 2% → 22% within 6 weeks |
| Q4 homepage revenue vs. prior quarter | +10% |
| Q4 homepage revenue vs. prior year Q4 | +18% |
| 2023 OKR outcome | Exceeded expectations across all owned KRs |
What I Learned
Speed is a strategic choice.
The decision to ship as H5 rather than waiting for native integration was the decision that made everything else possible. It compressed our learning cycle from quarters to weeks. In a market where AI trends have a shelf life measured in months, that compression was the difference between being early and being irrelevant.
Trends are readable; audience fit requires ongoing work.
We could see which aesthetics were gaining traction from competitor moves and social signals. The harder, continuous job was matching those trends to our specific user base across markets — not assuming that what worked in Japan would work in the US. Building the data infrastructure to make those calls quickly was what turned a one-time launch into a repeatable growth engine.
Channel economics aren't always visible in the headline metric.
Pop-ups had 22× the click-through rate of banners, and looked like the obvious winner. Banner-driven users converted to purchase at a dramatically higher rate. The learning generalises: volume and value are not the same metric, and optimising for one without understanding the other is a reliable way to misallocate budget.
A new monetisation model needs to create value, not just extract it.
The credits model worked because it gave users a genuine choice — pay for exactly what you want, without committing to a subscription. The result was incremental revenue and improved subscription conversion, rather than the revenue cannibalisation we were trying to avoid.
All metrics and performance data referenced in this case study reflect the 2023–2024 period. Sensitive figures have been anonymised or aggregated, and no proprietary or commercially confidential information is disclosed.