Campaign Success Requires More Than Predictions

Seeing The Future 

Not all conversions are created equal; some hold significantly greater value. Imagine two users—one signs up for a trial, while the other becomes a paid subscriber. It’s clear which user brings more value, but the real trick lies in spotting them early and fine-tuning your campaigns to attract more like them.

In value-based campaigns on platforms like Google, Meta, or TikTok, the focus shifts from merely increasing conversion numbers to striking the right balance between Lifetime Value (LTV) and cost, all to enhance Return on Ad Spend (ROAS) and maximize overall conversion value. The ad network AI, in theory, should identify which customers are most valuable—those who subscribe, upgrade, reorder, or refer others—and focus on acquiring similar users. But here’s where things get a bit tangled: most ad networks operate with attribution windows that range from a mere day to 28 or 30 days, while for many businesses—like SaaS PLG, marketplace delivery, fintech, or wellness—the actual conversion window stretches far beyond that.

‍

This creates a deeper conundrum because ad networks rely on a wealth of early signals to optimize campaigns. These signals close the feedback loop, identifying high-value users and finding more like them. But what if your conversion events occur well past the 7-day mark—at day 30, 60, or even later? Or what if your conversion volume is too small to provide enough data for the algorithms to chew on? This leaves the AI stumbling in the dark, blind to critical signals.

Offline Conversions, When It’s Not Enough 

To bridge this gap, many networks suggest importing conversions through offline APIs like Google Ads API, Meta’s Conversion API (CAPI), or TikTok’s Events API. But here’s the rub: businesses often stumble here, as forecasting future value is no walk in the park—especially with scant early data or when there’s a mismatch between early conversions and long-term user value.

Recognizing this dilemma, we set out to help businesses fine-tune ad network algorithms by crafting precise forecasts. In the early days of Voyantis, our approach was straightforward: whip up LTV predictions, send them to Google or Meta via server API, and cross our fingers. It worked—about 50% of the time. But after thousands of experiments, it became abundantly clear that the process was far more complex than we first thought. Predictions alone weren’t cutting it. The real challenge lies in weaving these predictions into the ever-changing tapestry of real-time data and navigating the unique, ever-evolving algorithms of each ad network—algorithms that seem to shift with every passing month. It’s about adjusting to dynamic conditions, tackling unseen variables, and making sure everything works together harmoniously to effectively influence ad networks.

It’s More Than Just Predictions 

To truly master value-based optimization, several factors beyond predicted LTV must be considered. How precise does your prediction need to be before sending a signal? How do you manage extreme values, like the occasional whale? Is there a backup model on standby in case data drifts off course? What’s your match rate? These are the crucial questions that determine whether the predictive signals you send will successfully nudge the ad network’s algorithm.

And then there’s the matter of timing. Ad networks thrive on early signals to fine-tune their algorithms, but there’s a tradeoff: sending early signals can lead to faster optimization but with less certainty, while waiting for more data may improve accuracy but delay action. Sometimes, this means updating signals as your confidence in the prediction grows. But beware—Google and Meta play by different rules: Google allows both upward and downward adjustments to predictions, while Meta only permits upward tweaks. Navigating this timing tradeoff across multiple ad networks can be as tricky as a rabbit in a hat.

Handling outliers is yet another hurdle. Imagine an accounting SaaS business where most users have LTVs between $200 and $700, but occasionally, you encounter users with LTVs of $5,000 or more. The challenge is ensuring the network recognizes these high-value users without letting such extreme figures muddle the algorithm and skew overall performance.

To tackle this, we poured countless hours into decoding ad network algorithms, figuring out how to send predictions at just the right moment and in a format their AI systems could comprehend and act upon. The solution we manifested has three key ingredients:

• A prediction layer to generate accurate LTV forecasts or custom predictions tailored to each business’s specific goals. For some, it might be all about the likelihood to subscribe, while others might focus on reorder probability.
• A prescription layer to blend and translate these predictive models into actionable signals that effectively influence ad network algorithms.
• A model continuity layer to ensure that predictive and prescriptive signals keep flowing smoothly to ad networks, even when data hiccups, schemes shift, or the wind changes direction.

How It All Works Together

With these ingredients in place, we built an AI orchestration layer that manages the entire process from start to finish, ensuring that your business attracts the most relevant customers. This layer employs an ensemble model approach, juggling hundreds of sub-models across dimensions like geography, ad network, segment type, and behavioral triggers. It constantly adapts in real-time, refreshing frequently—initially on an hourly basis, then switching to trigger-based updates as user data accumulates. This allows us to handle complex data patterns and unseen scenarios, keeping early signals sharp while maintaining long-term model performance. Specialized models for different stages of the user journey ensure accuracy every step of the way.

But we don’t stop at predictions; we transform those insights into actionable steps. These actions, driven by each user’s predicted future value, are seamlessly delivered across your platforms via APIs and integrations, automating the customer journey and driving exponential growth. In essence, we’ve mastered the art of making Google, Meta, and TikTok’s ad algorithms work in your favor.

As data shifts, business logic evolves, or goals change, our fallback models and self-learning systems ensure your campaigns remain optimized with the most accurate model outputs. Additionally, our prescriptive AI layer adapts alongside the ever-evolving algorithms of Google and Meta, ensuring your strategies stay ahead of the curve.

To wrap it up, success in value-based optimization goes beyond simply sending accurate LTV predictions. It’s about strategically influencing the ad network’s delivery algorithms. By continuously refining our models, balancing accuracy with timing, and adapting to the unique quirks of different channels, we’ve developed a robust system that maximizes performance and delivers impactful results.

‍