AI has quickly become an integral part of ad tech solutions. For publishers, however, the key question is not whether a platform uses AI, but what AI can actually do, what its decisions are based on and whether it genuinely delivers better results. In ad monetisation, individual decisions cannot be considered in isolation.
A change in one part of the auction can affect demand behaviour, ad delivery and ultimately publisher revenue. Making better decisions therefore requires not only looking at individual metrics, but also understanding how the different parts of the monetisation process affect one another.
This was also the starting point for our own AI development: what matters is not the model itself, but the data and ad tech expertise behind it, an understanding of each publisher’s own environment and the ability to demonstrate the value AI creates.
Start with the Data, Not the Model
AI decisions are only as good as the data they are based on. In ad monetisation, auction data alone is not enough. The full picture comes from combining information about auctions, delivered impressions and revenue with the context in which the auction takes place. Data granularity also matters, as demand behaviour and placement performance can vary significantly depending on the device, market, format or placement, even within the same publisher. This means that a single generic model may not work for every publisher. Models need to learn from each publisher’s own data and account for differences in placement, device and the broader auction context when making decisions.
A concrete example is Relevant Yield’s machine learning-based floor price optimisation, which is currently being actively tested ahead of its official release. Effective optimisation cannot be based on CPM alone. It also needs to account for bid participation, auction pressure, fill rate and the value of bids lost because the floor is set too high. Optimisation also needs to account for its impact across the entire auction chain, from Prebid to the ad server. Machine learning becomes valuable when sufficient data enables it to make better decisions than predefined static rules.
From Machine Learning to Agentic AI
Machine learning and large language models have different but complementary roles in ad monetisation. Machine learning is well suited to quantitative optimisation and learning from data, while LLMs can help analyse reports, explain findings and support user decision-making.
Relevant AI, launched in early 2025, uses language models within Relevant Yield, allowing users to query reports, interpret data and work with Relevant Yield’s product knowledge. We are now expanding Relevant AI in a more agentic direction. The upcoming AI Brief is one example: users can schedule the AI assistant to generate briefs at defined intervals, after which the assistant independently builds the brief, identifies and classifies relevant observations, and highlights areas that may require the user’s attention.
In the future, the same agentic approach can be extended by giving the AI assistant additional instructions and tasks. As automation advances, it becomes increasingly important for publishers to retain control over what AI can access, which actions it can take automatically and which require approval. Any changes made should also be observable, traceable and reversible.
How Do You Separate Real Value from Hype?
Publishers should look at what an AI system’s decisions are based on: what data it can use, whether it can connect auction data with actual performance and revenue, and the level of granularity at which decisions are made. Equally important is the ability to verify their impact. In optimisation, for example, results should be compared against a control and examined in greater detail rather than only at an overall level..
Ultimately, the real value of AI in ad monetisation does not depend on any single model, but on the data, expertise and technology around it. As AI expands from analysis into optimisation and agentic capabilities, questions about what AI can do automatically and how much control remains with the publisher become increasingly important.