How AI App Studio 44pixels Increased ROAS 23% by Predicting Subscriber Value From the First Session
By the numbers
Overview
Company
44pixels is the company behind the successful AI-powered apps Pixi, Cue, a hands-free meeting notetaker, and Clara, the AI that answers the phone.
Industry
Mobile Apps
Headquarters
London, United Kingdom
Campaign Type
Google Search
Most Valuable Users
Annual subscribers and weekly subscribers who remain active past their first month
Snapshot
The real growth advantage is not acquiring more subscribers at any cost; it is identifying the right subscribers early enough to influence the auction. Voyantis helps us make long-term customer value an actionable acquisition signal.
The GenAI consumer app market is won and lost on acquisition efficiency. In a fast-moving category, the ability to identify and scale high-value users quickly creates a meaningful competitive advantage. What decides who wins is finding and keeping the right subscribers.
App studio 44pixels builds and scales AI-powered applications across consumer productivity and small-business use cases. Cue, its AI-notetaking app, runs on weekly and annual subscriptions – two tiers with very different value profiles. Running tCPA against subscription starts meant paying the same to acquire every user, regardless of their value.
44pixels partnered with Voyantis to translate its subscriber and product engagement data into predictive signals that Google could optimize against.
Impact at a glance
Increased ROAS 23% and reduced cost per subscriber by 6%
with each subscriber across tiers priced according to their predicted value from the first session
Shifted the subscriber mix toward high-intent users
by replacing a flat conversion event with predictive signals cutting refund rates by 33% and cancel rates by 12%
Engineered signals tailored to Cue's multi-tier subscription funnel
with predictions timed to filter early churn and calibrated by market
Laid the foundation for full-account expansion
with the US rolling out and all remaining campaigns to follow
The challenge
tCPA optimization was effective at driving subscription starts, but it could not distinguish between subscribers with very different retention and refund profiles. Shifting to value-based bidding was the obvious next move, but Cue's subscription structure introduced a set of trade-offs that couldn't be solved with a single model or simple bidding shift.
The value difference between weekly and annual subscriptions is significant at any early time horizon. Optimizing too aggressively toward annual subscribers quickly produced diminishing returns, as that audience was smaller and more expensive to acquire. Shifting weight toward weeklies solves the volume problem but trades it for another – a broader weekly-subscriber population with more varied retention profiles.
The solve
Voyantis combined strong predictive modeling with the practical signal engineering required to make those predictions useful inside Google's bidding system. Their team worked closely with ours to build models around our actual subscription, product-usage and acquisition data, rather than applying a generic approach.
Model each subscriber's predicted value through two full renewal cycles
Weekly and annual subscribers churn for different reasons, at different rates, on different timelines. Voyantis designed one set of models for each rather than training on a blended population that would average out the differences between them:
- Annual Renewal Model – Will this yearly trial subscriber renew or cancel?
- Weekly Retention Model – How many weeks will this weekly subscriber retain?
Each model produced a dollar LTV value calibrated to how that subscriber group actually behaves, with predictions projected far enough to capture two full annual renewal cycles and the long-tail retention of weekly subscribers who stay.
Each model was trained on a broad set of privacy-safe contextual and behavioural signals available early in the subscriber journey, including:
- Early Product Engagement: Whether the user engaged with core product features in their first session.
- Subscription Depth: How far the user progressed through the subscription flow and whether they took action.
- Acquisition Context: Which channel and market the user arrived from.
Engineer subscriber value into a signal Google can optimize against
Voyantis translated each model's output into a calibrated signal Google could bid against, because a prediction means nothing to Google until it carries a dollar value.
- Timing: Signals were timed to each subscriber type's cancellation risk window. Subscribers who converted quickly received a high-confidence signal once that window closed. Those still in trial were tracked on a longer horizon, with predictions updated as behavioral data accumulated rather than forcing a call on incomplete information.
- Value Calibration: Subscriber revenue varies significantly by market, and bidding as if it didn't would misprice every auction. Voyantis calibrated the signal to Cue's actual revenue by market so each bid reflected what a subscriber in that region was worth.
Activate the signal in a live auction
Cue's web-to-app campaign structure meant a user's journey from ad click to conversion crossed three separate systems, each tracking identity differently. Before any prediction could reach Google, Voyantis built the matching infrastructure to bridge those handoffs reliably.
Together, the teams structured a controlled test across 44pixels' largest markets – a tCPA control arm optimized against trial starts running against a tROAS test arm optimized against Voyantis's predictive signals. Results were measured against subscriber quality metrics: refund rates, cancel rates, and ROAS at multiple time horizons.
Learn and evolve as traffic shifts
To keep signals sharp and pointed toward the right mix, Voyantis tracked and maintained three layers:
- Pipeline integrity: Data freshness tracked across every upstream source, with breaks flagged before they could delay predictions.
- Model Health: When over-prediction emerged in a specific segment, Voyantis isolated it and recalibrated without touching what was already performing.
- Learning Preservation: As predictions sharpened, Google's accumulated signal compounded rather than reset.
Voyantis’ impact
23%
increase in ROAS
33%
decrease in refund rate
12%
decrease in cancel rate
6%
decrease in cost per subscriber
Higher-intent subscribers at a lower acquisition cost
Across 44pixels' major markets, the subscriber mix shifted toward higher-intent users, with each acquisition priced according to predicted value from the first session. The results:
- 23% increase in ROAS
- 33% decrease in refund rate
- 12% decrease in cancel rate
- 6% decrease in cost per subscriber
Refund and cancel rate improvements reflect a structural change, confirming a shift in who is actually coming through the door. A core predictive infrastructure is now wired into 44pixels' growth strategy, with the signals tuned to each app's own subscriber dynamics. The infrastructure gives 44pixels a repeatable foundation for applying predictive acquisition across future products.
Voyantis has earned its place in our growth stack. Together, we turned our subscription and product-engagement data into an acquisition signal that now informs how we bid, where we invest, and how we approach the launch of new products.
