Growth & CRO
AI image greetings lifted feature engagement by 168%
· Updated · 5 min read
S
Get CRO Checklist — 17 fixes
The 17-point checklist from the 73%→54% teardown. Same format, new numbers.
No spam, unsubscribe anytime. I only email teardowns.
The short answer
A dormant "Send Greetings" feature became a +168% engagement story when we fused Google's Nano Banana (Gemini Flash image) into it. The unlock was geo-segmentation: 27 festivals across five Indian regions, surfaced at the right moment in the merchant's existing workflow. The AI did the artwork; the calendar did the work.
The feature had been live for years. Average usage: 0.2 cards per active user per month — statistically, decoration. Eighteen months later the same feature was at 3.4 cards per user, engagement up 168%, and the thing that changed wasn't a redesign or a re-launch. Send Greetings in LiveKeeping PRO+ got two ingredients: an image model that could render Bengali script without tofu boxes, and a festival calendar that knew a Kolkata trader and a Pune shopkeeper don't celebrate the same week.
Why it sat unused
- The feature existed for years but had no reason to be opened.
- Every greeting-launched list was static; no "when."
- Nothing surprised the user into trying the generator.
- The templates looked like every other business's spam — same Diwali image, same generic "Happy Diwali" text. No brand identity, no personalization.
The AI reinvention: model + calendar + timing
We integrated Google Gemini 3.1 Flash Image (Nano Banana) so users could build a festival card instantly — no design skills, no assets, no excuses.
But the model alone doesn't fix engagement. The timing inventory did:
- A geo-segmented festival calendar across five Indian regions — Pan-India, South India, Maharashtra/Goa, East India, and Gujarat.
- 27 occasions — Diwali, Durga Puja, Ganesh Chaturthi, Bihu, Vishwakarma Puja, Gudi Padwa, Ratha Yatra, Onam, Pongal, and 18 more — each assigned the right messaging, language, and visual template.
- Triggers placed where users already worked — send-time reminders, invoice completions, payment confirmations, GST filing completions.
The three-part build
| Part | What it does | Why it moved the needle |
|---|---|---|
| AI Greeting Generator | Select occasion → business name + customer name (pre-filled from ledger) → language → Nano Banana generates a custom card in 15 seconds. No design skills. | The AI was the paint. The user never touched a design tool. |
| Geo-Segmented Festival Calendar | 27 occasions Mar–Sep 2026 across 5 regions. A West Bengal trader sees Ratha Yatra and Vishwakarma Puja. A Maharashtra shopkeeper sees Ganesh Chaturthi and Gudi Padwa. | No national blast. The right festival reaches the right user. |
| Evergreen Greeting Library | 13 on-demand types any day — Good Morning, Good Night, Thank You (post-payment + post-order), Motivation (Monday + general), Payment Reminder (3 urgency tiers), Sale Promotion, 4 discount tiers. | Usage didn't wait for a festival. Daily habits = daily engagement. |
The model: Nano Banana (Gemini 3.1 Flash Image)
We chose Nano Banana for three reasons:
- Precise text rendering on images — Indian scripts (Devanagari, Bengali, Tamil, Gujarati, Telugu) render cleanly on greeting cards. Most open models garble non-Latin scripts.
- Multi-language / regional script support — The model understands "शुभ दीपावली" and renders it correctly in the card, not as tofu boxes.
- Greeting card & marketing asset generation — It was trained for this exact use case: marketing mockups, greeting cards, branded content with in-image text translation.
The integration: a thin wrapper around the Gemini image generation API, cached responses for repeated occasion+language combos, fallback to a pre-rendered template if the API fails. Total integration: ~400 lines of TypeScript.
The calendar: 27 occasions, 5 regions, zero national blasts
| Region | Key occasions (sample) | Language priority |
|---|---|---|
| Pan-India | Diwali, Holi, Eid, Christmas, Republic Day, Independence Day, Gandhi Jayanti | Hindi, English |
| Maharashtra/Goa | Ganesh Chaturthi, Gudi Padwa, Narali Purnima, Vat Purnima | Marathi |
| South India | Pongal, Onam, Ugadi, Vishu, Thiruvonam, Karthigai | Tamil, Telugu, Kannada, Malayalam |
| East India | Durga Puja, Ratha Yatra, Vishwakarma Puja, Kali Puja, Poila Boishakh | Bengali, Odia, Assamese |
| Gujarat | Navratri, Uttarayan, Janmashtami, Diwali (regional), Bestu Varas | Gujarati |
Each occasion has: a default prompt template (business name, customer name, occasion, language), a visual style guide (color palette, motif), and a trigger date range (e.g., "Diwali: 3 days before through day of").
The triggers: where the user already is
We didn't build a "greetings tab" and hope they'd visit. We injected the generator into existing workflows:
- Invoice completion → "Wish your customer a happy [upcoming festival]?"
- GST filing success → "Your return is filed. Wish your clients a prosperous [festival]?"
- Payment received → "Thank your customer with a [festival] card?"
- Send-time reminder — 9 AM on festival morning: "It's [festival] today. Send wishes?"
The greeting door was the notification banner — not a separate tab.
The A/B tests that shaped the product
We didn't ship the final version on day one. We tested:
| Test | Variant A | Variant B | Winner | Lift |
|---|---|---|---|---|
| Greeting door | Separate tab | Notification banner | Banner | +340% CTR |
| Prompt style | Free text | Occasion picker + prefill | Picker | +220% completion |
| Image model | Stable Diffusion | Nano Banana | Nano Banana | +180% quality rating |
| Timing | Morning only | Morning + invoice | Both | +45% daily active |
The notification banner was the single biggest win. Users don't navigate to features — features must navigate to users.
The prompt engineering (what actually worked)
The prompt template that produced the best cards:
Create a festive greeting card for {occasion}.
Business: {business_name} ({business_type})
Customer: {customer_name}
Language: {language}
Style: {style_guide}
Constraints:
- Include business name prominently
- Include customer name
- Render greeting text in {language} script correctly
- Color palette: {color_palette}
- Aspect ratio: 4:5 (portrait card)
- No watermarks, no extra text
We cached responses by (occasion, language, style_guide) — 80% of requests hit the cache after week 2. Cache TTL: 7 days.
The numbers
| Metric | Before | After |
|---|---|---|
| Feature engagement | dormant | +168% |
| Occasions surfaced | ~1 national | 27 geo-segmented |
| Regions covered | 1 | 5 |
| Platform users | 50,000+ LiveKeeping SMBs | same |
| Avg cards sent / active user / month | 0.2 | 3.4 |
Users didn't adopt an AI generator. They adopted a card for the occasion before them — the AI was the paint, the timing was the purchase.
What I'd do differently
- Air the calendar earlier. The 5-region mapping should've shipped before we raised the model; geo-accuracy was the real feature.
- Test the greeting door. Our best surface was a banner inside a richer notification — I'd A/B it against "merchant made a bill → invite to wish a client" earlier.
- Build the evergreen library first. The daily-use greetings (Good Morning, Thank You, Payment Reminder) drove 60% of post-launch volume. Festival cards are seasonal; evergreen is the retention engine.
The full build is in the Send Greetings + Nano Banana case study. It pairs well with the push notification architecture this automated: timing and geo are 90% of the lift.
Frequently asked questions
What makes an AI feature actually used?+
Crawl-through, not novelty: a one-line image generation fused into an existing habit (sending festival wishes) beat a lobby feature.
How do you geo-segment festival greetings?+
Five Indian regions each have their own festival calendar; we mapped 27 occasions so nothing sent a one-size national blast.
How was feature engagement measured?+
Engaged users as a share of PRO+/PRO before vs after the AI launch. Same login floor, different feature taps: +168%.