FAQ: Why Signal Engineering is The Missing Link in Predictive LTV Optimization

Real installs, real purchases, and real revenue. 

That’s what ad platforms have been training their delivery algorithms on for the past 15 or so years – but that’s not where the best growth teams are gaining an edge today. 

They’re focusing on predicted lifetime value (pLTV). 

pLTV is the revenue – or value – a user is predicted to generate over a defined window of time, like 30, 90, or 180 days.

They care about the value a customer will create, not just the conversion they already completed. They recognize that all conversions are not created equal, so they’re optimizing for predicted revenue, predicted propensity, predicted retention, and even churn likelihood. 

But in order to do this, they need a way to translate their predictions about customer value into a language that ad platform algorithms – which have only been trained to understand historical, deterministic data – can actually understand and act on. 

Signal Engineering is that translator. This FAQ will explore how Signal Engineering is the missing link that allows growth teams to optimize their campaigns for pLTV and focus on their highest value customers. 

What Is Signal Engineering?

‍Before we go any further, let's define exactly what we mean when we say Signal Engineering. 

Signal Engineering is the structured, dynamic process of translating predictions into signals that are intelligible, timely, and impactful inside ad platforms. It accounts for the platform’s learning logic, your user journey, and the nature of the prediction so bidding and delivery move toward tomorrow’s high-value users and away from yesterday’s cheap conversions.

What do I need in order to optimize my paid campaigns for pLTV? 

In order to optimize paid campaigns for pLTV, two core capabilities are required. You need a strong and resilient prediction engine and a sophisticated Signal Engineering process. 

A Strong and Resilient Prediction Engine

You need a prediction engine that can: 

• Forecast a user’s future revenue or behavior early, ideally within just hours of app install or site visit. 
• Update continuously and always be available. 
• Be resilient to drift, downtime, missing data or specific data pipeline issues. 
• Allow predictions to evolve over time as the user interacts with your product.
• Generate predictions that are accurate early enough to be used within the ad platform's conversion window (oftentimes that’s limited to 7 days). Both Meta and Google give more weight to conversions that happen soon after the ad interaction.

‍

A Sophisticated Signal Engineering Process

When dealing with deterministic data, sending a signal as an event via API is easy – it’s what we’ve all been doing for years. 

But when it comes to predicted events, ad platforms provide no best practices, documentation, or clear guidelines for turning your predictions into a signal that they can understand. That leaves growth teams asking questions like: 

• Should we represent a user’s value through the volume of signals, or through the event’s value? 
• Which of those options performs better? 
• Should the signal be sent very early – risking lower prediction accuracy – or wait for higher certainty at the cost of timing? 
• What would the network’s algorithm prefer? 

This list keeps going, but you see the point – there are a lot of important questions that have to be answered in order to successfully implement Signal Engineering. 

If those questions aren’t answered, and you don’t implement a Signal Engineering process, your predictions are just that – predictions. They may likely not drive the impact you’d expect in the ad platforms delivery algorithm and leave you unable to optimize your paid campaigns for pLTV. 

How do I implement Signal Engineering? 

There are four key areas of focus when it comes to Signal Engineering: Timing, Cadence, Value, and Strategy. Let's dig into each of them.

1. Timing

The thing to consider: When should the signal be sent?

The risks: Send it too early and it may be inaccurate. Send it too late and it loses effectiveness. 

How to engineer it: Signal Engineering models the maturity curve of each prediction and aligns it with platform preferences. Google, for example, prioritizes early signals, while Meta prioritizes accuracy. Google allows for both increases and decreases when updating conversion values, while Meta only accepts increases. This makes Meta’s delivery algorithm more sensitive to accuracy – even at the expense of timeliness.

2. Cadence

The thing to consider: How often should the signal be sent?

The risks: Send too many updates and you’ll flood the ad platform and cause confusion. Send too few updates and the platform might deprioritize your signal because you’re not providing enough to train their delivery. 

How to engineer it: Cadence must reflect signal confidence, value volatility, and platform thresholds. It should also account for the overall signal volume per ad group or campaign, which can influence optimization behavior on the platform.  

3. Value

The thing to consider: What value should you send, and how should it be framed?

The risks: With predictions, this gets trickier. In a predicted LTV model – where each user has a future dollar value – the spread is often much wider than actual D7 revenue. That means more edge cases and outliers, and a single extreme value can throw a campaign off. On the other hand, not all predictions represent dollar values. Propensity models output probabilities between 0 and 1, and platforms typically don’t respond well to very small numbers.

How to engineer it: Normalize before you send your signals. For pLTV models, you can transform raw dollar predictions into rank-based scores and use logarithmic or exponential scaling to reduce the influence of outliers. For propensity models, multiplying scores by a constant can help them register, whether that constant is the average LTV, a custom exponent, or a learned scaling function. Each case requires careful design, and the strategy should be tailored per platform, per use case, and per prediction type.

4. Strategy

The thing to consider: How do you translate your business goals into a signal the platform can optimize toward?

The risks: Choosing the wrong format for your objective or platform can cause the campaign to underperform.

How to engineer it: Signal strategy determines the format and design of the signal itself.

• Send a single event with a value when optimizing with value-aware objectives like tROAS, pMAX, or VO.
• Express user value through signal volume for CPA-optimized campaigns or where value optimization is less stable (e.g., a high-quality user might trigger ~30 signals whereas a low-quality user might trigger just one signal).
• Use thresholding to emphasize acquisition of high-value users, sending signals only for users predicted to be good or great.
• Choose the representation intentionally. For example, a predicted $100 can be sent as a single $100 event, five $20 events, or 100 micro-events at $1 each. Each approach influences how the algorithm learns and adapts.

‍

What happens if I don’t use Signal Engineering? 

You can’t just send a raw prediction to an ad platform and expect results.

For example, if your model predicts a user is worth $112.37 in the next 180 days, you might be tempted to send that number as a conversion value to Google or Meta.

But this is where most pLTV strategies fail.

Why? Because platforms are not built to interpret predicted outcomes. Their delivery algorithms were trained across hundreds of billions of ad impressions to react to deterministic signals, such as:

• Events that actually occurred (like purchases)
• Values that actually materialized (like order revenue)
• Signals that have clear timestamps and conversion windows

Predictions are an entirely different animal. 

They're probabilistic. They evolve. They represent behavior that hasn’t happened yet. Most critically – they’re not events.

Sending predicted revenue to a platform designed to interpret deterministic outcomes can lead to:

• Misinterpretation of the value
• Campaign instability
• Erratic learning behaviors
• And wasted spend due to misaligned optimization

Predictions without translation are noise. Signal Engineering turns predictions into event-shaped, timely feedback the platforms can actually learn from.

Still have questions? 

If you’re wrestling with how to make pLTV legible to a specific platform, ping me – this is exactly what we help teams with at Voyantis.

Itai leads Product Growth at Voyantis, where he helps advertisers improve performance by optimizing the signals ad platforms learn from. In ad‑tech since 2013, his career spans automated bidding, creative intelligence, and platform‑side signals, most recently leading App Signals product strategy at TikTok, giving him a front‑row view of how automation is reshaping growth. Today, he focuses on pairing strong prediction engines with rigorous Signal Engineering to make pLTV actionable at scale.