Optimized to Matter: The Case for Predictive ML in an Age of GenAI

By Larry Price, PhD, Head of Data Science, Koddi

In digital advertising, relevance is easy to talk about but notoriously hard to execute. It’s the difference between an ad that feels helpful and one that feels like noise — and it’s a major driver of performance across the entire ecosystem.

According to IAB Research, almost 90% of consumers prefer personalized ads and 87% are more likely to click on ads for products they’re interested in or shopping for.

Relevance is one of the most misunderstood and underused levers in digital advertising. It’s hard to define, harder to measure, and often overlooked in favor of flashier trends.

Make no mistake: generative AI is absolutely game-changing technology. It’s reshaping many aspects of the industry — from how creative is produced to how content is generated and interfaces are designed. But when it comes to deciding which ad to show, to whom, and when, traditional predictive machine learning still does the heavy lifting. Generative AI will keep redefining how ads are made — but predictive ML currently determines which ads get seen. That’s where ad relevance lives.

What Makes an Ad Relevant?

Relevance is context-sensitive. If I’m searching for tacos and see an ad for pizza, that might feel off. But if the real intent is “something quick for dinner,” it might be spot-on. That nuance—understanding the latent intent behind observable behavior—is where ML shines.

Modern ML models can learn from browsing patterns, timing, device signals, historical interaction data, and more. The challenge is integrating all of that in real time to make a useful prediction, not just a precise one. That’s why the best systems now optimize not just for clicks or conversions, but for business-aware relevance—accounting for margin, inventory constraints, and downstream outcomes.

This is where experimentation matters. You don’t learn what “relevance” really means in your ecosystem by intuition or committee. You learn it by testing hypotheses, shipping changes, and letting the market tell you what works. ML makes that loop faster and smarter.

Aligning Incentives Across the Marketplace

Relevance is only sustainable if it works for everyone in the system: the advertiser, the publisher or retailer, and the consumer. Advertisers want return on ad spend. Publishers want yield and a quality user experience. Consumers want ads that don’t feel like ads.

The best machine learning systems don’t just optimize for one of those—they help mediate tradeoffs between all three. For instance, models can forecast expected ROAS across different customer segments and auto-adjust bidding strategies accordingly. That leads to more efficient spend, better user experiences, and more monetizable moments.

But this only works if everyone is clear on the objective function. Too often, relevance is framed as a fuzzy concept when what we really need is sharper definitions, better metrics, and more data transparency between partners.

The (Still Underestimated) Power of First-Party Data

If relevance is the outcome, first-party data is the foundation. It’s the most durable, privacy-safe signal we have—and it’s increasingly central to how retailers and platforms differentiate themselves.

Used properly, first-party data helps personalize without overstepping. It enables better targeting, smarter measurement, and more accurate predictions. But it also comes with responsibility. Every ad system needs to respect user privacy, align with regulation, and earn consumer trust. That means being transparent about what’s collected and how it’s used, and designing systems that work even when data is sparse.

Building for Real-Time Relevance

Ad relevance isn’t static. It shifts with user behavior, seasonality, supply and demand, and countless other factors. So our models—and our infrastructure—need to evolve with it.

A robust relevance system should:

• Score ads in real time for every auction
• Refresh frequently to adapt to market dynamics
• Support modular experimentation for model variants, objectives, and ranking logic

Yes, this requires real investment in infra—model serving, feedback loops, observability—but it pays off in performance, agility, and strategic control.

Don’t Let the Spotlight Fool You

Generative AI is having a well-earned moment. It’s opening up new possibilities and accelerating how we build, test, and deliver content. Behind the scenes, predictive machine learning remains essential — especially when it comes to decisions about which ads get shown, to whom, and when.

As the industry evolves, the most effective platforms won’t choose between generative and predictive models — they’ll find the right role for each.