Mastering Campaign Optimization: FAQs on Predictive AI and Scaling Strategies

Using advanced techniques like predictive modeling to enhance campaign performance alongside Google’s own AI algorithms is both complex and nuanced. Success requires a clear understanding of how these models can drive better results—and when additional support might be needed. Even if your team is leveraging revenue data or experimenting with predictive models, the outcomes may still fall short of expectations. Let’s dive into what might be missing and how you can optimize your approach.

Why can't I just send my actual revenue figures? 

Google prioritizes early signals generated shortly after a user interacts with an ad. However, in the first 7 days post-ad interaction, users rarely complete monetization events like purchases or subscriptions. Instead, actions like email signups or questionnaire completions are more common. This makes predicting long-term metrics like LTV or 30-60 day ARPU challenging. What’s more, optimizing for D7 ROAS using short-term data often yields mediocre results or even confuses Google’s algorithm.

‍“You need early signals, but rarely do you have a clear understanding of the monetization events that truly impact Google within the first 7 days.”

The better approach is to use a predictive model that looks beyond top-funnel events to long-term user-level revenue. Voyantis updates predictions looking at deep funnel events within hours of first ad engagement and sends only the most informative signals to Google, aligning your campaigns with long-term growth objectives.

I can send static values. Do I need to do anything else?

“User journeys are dynamic by nature; often, a user’s true value becomes clear only over time.”

Static values work well for dimensions like keywords or geographic targeting, as seen in tCPA campaigns. While tROAS can also use static values, dynamic attributes like user interactions and evolving behaviors require a more advanced approach. Predictive AI models adjust in real-time based on changing factors, creating weighted outputs that precisely predict user value. This ensures your campaigns align with real-time data for maximum efficiency.

‍
Your team is building predictive models in-house. Shouldn't that be enough? 

“Predictive models alone aren’t enough—you need signals delivered quickly and in a format that Google can learn from and act on effectively.”

In-house predictive models are valuable for scenarios where instant optimization isn’t necessary. However, these models often take over a week to deliver predictions. Google needs signals within 24 hours of user interaction to optimize effectively. Voyantis generates user-level predictions within hours, providing timely signals for Google’s algorithm. Your in-house models still play an important role in post-campaign analysis and future strategy refinement.

Will Value-Based-Bidding (VBB) hurt my scale?

“Focusing on deep funnel events helps you target higher-value customers while rarely affecting FTB volume.”

Scaling requires casting a wide net, but not all acquired users become repeat buyers, making scale alone insufficient ; you don’t want to waste resources. Still you also don't want to see a drop. While CPA for upper-funnel events may fluctuate slightly, key metrics like First-Time Buyer (FTB) are rarely negatively affected. At Voyantis, we closely monitor campaign performance, and when necessary, maintain scale by combining FTB and LTV signals. This approach ensures you can effectively balance growth with efficiency.

Why Optimize for LTV / Second Purchase?

You’re already accounting for value variations by setting separate tCPA values for different ad groups. However, this approach misses significant customer differences, such as engagement patterns or device usage. Leveraging these differences can improve efficiency. Google campaigns optimized this way have seen ROAS gains of over 20%.

So what should you do? 

“Voyantis’ AI prescription layer determines which signals to send to Google and when, all at the user level.”

Create predictions for each signup, focusing on events like 60-day LTV or second-purchase probability. Update predictions within the first 7 days based on user activity, including FTBs. Voyantis’ AI prescription layer decides which signals to send to Google and when at the user level. For instance, only predictions for non-buyers with a purchase probability above a certain threshold might be sent.

Why not predict the probability for a First Time Buyer?

The network doesn’t need a prediction for FTB because it already has immediate visibility into these users through early signals. Google’s algorithm prefers prompt, real-time signals rather than predictions for users higher up the funnel, such as signups.

In our experience, trying to predict for a population that is farther up the funnel from your current optimization goal (e.g., FTB) does not lead to better performance. While it might seem intuitive that providing more signals would help, it often fails to compensate for the inefficiencies introduced by moving up the funnel. A predictive model for FTB is unlikely to outperform a campaign directly optimized for FTB.

What does a good test set up look like? 

✅  To set up a strong test, use Google Drafts and Experiments. 

✅  Select one to three large campaigns and consolidate them into a single campaign.

✅   Run a 50/50 A/B test for eight to twelve weeks. 

✅  Use your existing BAU tCPA goal as the control, and optimize the test for tROAS using Voyantis’ predictive signal.

We hope these insights help you and your team evaluate the best setup for your campaigns and weigh the benefits of using static or in-house predictive models versus partnering with a solution like Voyantis. By not only creating dynamic models but also delivering them in a format and timing that ad networks can effectively learn from, Voyantis can take your data-driven optimization to the next level.

‍

‍