How Credit Genie Grew ARPU by Replacing Conversions with Predictive LTV Signals
By the numbers
Overview
Company
Credit Genie is a financial app
Industry
Fintech
Campaign Type
Google App Campaign
Snapshot
Value-based bidding was the natural next step for us, but getting it right meant more than plugging in a predictive model. Voyantis brought the team and technology to tip the auction in our favor, bringing in higher-value users.
Credit Genie had built a strong acquisition engine on Google App campaigns, optimizing around an early, fast-firing conversion event that directly drove revenue. This strategy brought in volume, but the growth team knew their best users were the ones who stuck around and kept coming back, not just those who converted once. Optimizing for that early action taught Google to find more people who were likely to take it, regardless of whether they would stick around long-term.
Credit Genie set their sights on bringing long-term value inside the learning window, but doing it right required predictive signal infrastructure they didn't have in-house. They turned to Voyantis to bring future value into the auction.
Impact at a glance
Grew ARPU 22% and ROAS 10% against Credit Genie's control tCPA setup
shifting the mix toward users who generate more revenue over time, not just more conversions
Kept the predictive signal accurate as the business and auction evolved
monitoring data pipelines and retraining the model as the product and user mix shifted
Built a proven foundation for predictive value-based bidding
shaping how Credit Genie acquires their customers at scale
The challenge
The gap between Credit Genie's high-value users and low-value ones showed up over months of repeat engagement and renewed revenue. However, Google's algorithm does most of its learning in the first several days after an ad click – nowhere near enough time to see who would turn out to be long-term subscribers. Most of Credit Genie's users convert on the first offer on the day they arrive, so any useful prediction had to be made within hours of a user's first interaction with the app, from a thin trace of early behavior.
On top of that, "valuable" wasn't a fixed definition. Like any rapidly growing company, Credit Genie adjusted frequently to keep pace with its own model and product changes, and the early behaviors that marked an engaged customer changed along with it. A model trained on one quarter's cohorts would keep bidding confidently on the last quarter's definitions, judging users by rules that no longer applied.
Credit Genie needed a way to predict that value early, deliver it to Google in a format the auction could bid against, and keep that current as the business itself evolved.
The solve
Credit Genie doesn't stay still. We're always evolving our business and product, so whatever we build with has to move at the same speed we do. Voyantis worked in lockstep with us along the way, rigorously testing and adjusting alongside every change to ensure continuity.
Predict value early in the funnel
Voyantis started with a deep dive into Credit Genie's growth context, business model, revenue streams, and growth KPIs, to identify a clear target: Day 90 LTV, predicted shortly after a user signs up. Ninety days spans several billing cycles, long enough to see whether a user stays subscribed.
No single model feature predicts value on its own. It's the combination of dozens of insights that allows Voyantis's model to form a confident read on a user's value within hours of their first interaction with the app, including:
- Session rhythm: how many sessions a user logged, how long they lasted, and how much time passed between them.
- Early user choices: even small decisions a user made early on carried real insight about long-term value.
- Disengagement signals: certain early drop-off behaviors before a user fully converted served as an early warning.
The model never sees revenue. Its prediction is sent to Google before the user has paid anything, so there's no payment history for it to look back on. It forecasts future revenue entirely from early behavior, and Credit Genie uses it solely to inform value-based bidding on Google.
The model kept prediction error under 3%, week over week. It was designed to distinguish users likely to stay engaged long-term from those who aren't, and to hold that distinction within each group, so Google could tell not just who was worth more, but by how much.
Engineer a bid-ready predictive signal
A prediction is not the same thing as a signal an auction can use. Turning Credit Genie's ninety-day value predictions into something Google could bid on required its own infrastructure.
Voyantis' automated signal engineering layer handled:
- Timing + Cadence: Not every user who signs up is ready to convert right away. Voyantis anchored its earliest signal just a few hours after sign-up, built from that first prediction. A single, well-timed prediction per user kept optimization focused.
- Value Calibration: Credit Genie's product structure meant predicted values clustered close together. Values were capped and floored to carve out real separation within that tight range, so Google could still tell high-engagement users from average ones. For users who hadn't yet converted, an early positive prediction and the signal built from it faded the longer that user went without taking action.
Test the signal in the auction against Credit Genie's own revenue
To confirm that a predictive signal could shift Credit Genie's user mix, Voyantis orchestrated a geo-split experiment on Android, running a tROAS campaign fed by signals of predicted ninety-day value against Credit Genie's existing tCPA setup. The test ran for eight weeks, with spend held close to perfect parity between the two groups throughout.
Success was read against Credit Genie's own numbers, not Google's reported conversions. That distinction matters because an ad platform can report a healthy conversion rate while still optimizing for users who churn within weeks – exactly the blind spot this test was built to close.
Keep the signal accurate as the business and auction shifts
The users a signal wins today become the data it learns from tomorrow so, left alone, it drifts. Catching these shifts depends on three things Voyantis runs continuously:
- Pipeline integrity: 24/7 monitoring of the data feeding predictions, catching a break before it impacts the signal.
- Model health: As cohorts mature, predictions are checked against real outcomes and Voyantis retrains the model to keep pace.
- Learning preservation: Voyantis rolls out updated models without resetting Google's accumulated learning, so improvements compound instead of restarting from scratch.
Voyantis’ impact
10%
increase in ROAS vs. tCPA
22%+
increase in ARPU vs. tCPA
A cleaner user mix that compounds in value over time
Tested against Credit Genie's original tCPA setup, the predictive signal delivered a 10% lift in ROAS by finding users more likely to stay engaged long-term. ARPU rose more than 22%, confirmed in Credit Genie's own revenue, within a point of what the model had predicted.
With that result proven, Credit Genie raised the bar and tested Voyantis's predictive signal against its own value-based bidding, which didn't incorporate long-term predicted value.
At double the original scale, Voyantis predictive signals delivered:
- 25% increase in ROAS
- 17% decrease in cost per valuable user (users who accepted a cash advance)
- 7% increase in ARPU
Two tests, two different bars to clear, and the same result each time: the predictive signals beat whatever it was measured against.
