Building the foundation for commerce media in the agentic era

The modern shopper journey is rarely a straight line.

 

Consumers bounce between retailers’ sites, social feeds, publisher pages, mobile apps and more. As they move across the commerce ecosystem, they leave behind a breadcrumb trail of signals that never quite match up.

Cookies in one corner, CRM IDs in another, and device identifiers off doing their own thing.

 

For marketers, this creates a couple of challenges. First, they have to deliver smooth, personalised experiences for consumers no matter where they shop. Second, they need to do this in an environment full of fragmented, mismatched signals.

 

This is where artificial intelligence steps in. Not the vague, hype-driven AI everyone talks about in a generic way, but a core capability that turns scattered commerce data into something you can actually use – all while laying the foundation for the agentic era ahead.

 

AI is already in the driving seat

 

Long before the AI buzz of today and the promise of the agentic future, artificial intelligence played a key role in commerce media. It’s driven bidding strategies, product recommendations and audience targeting, making media buying more efficient and more relevant.

 

What’s different today is the step-change in capability brought by deep learning and, in particular, generative AI. Models can now ingest far more data with intrinsically higher complexity – from images and videos to multi-lingual inputs – and tackle a wider range of tasks, from drafting marketing briefs to creating videos from still images. All of this raises the bar for what “good” looks like in performance marketing.

 

Our approach at Criteo goes beyond just integrating off-the-shelf AI models into our platform. Instead, our AI tools draw on a tailored mix of proprietary models developed at Criteo and select external models that cover capabilities we don’t consider essential. We’ve built our AI with commerce in mind, and our homegrown models are trained on nearly two decades of rich, commerce-focused signals.

 

Today, this commerce AI is embedded across our Commerce Media Platform, powering everything from predictive bidding to creative personalisation in a way that’s tuned to real shopper behaviour.

 

Built for commerce outcomes

 

What makes Criteo AI different isn’t just the in-house expertise we leverage to craft high-performing models – it’s also in the data itself.

 

Our commerce AI is trained on commerce signals. To be more specific, it’s trained on data from 4.5 billion SKUs, 3700 product categories, 17,000 advertisers and the shopping behaviour of around 720 million daily active users. The depth and granularity of this data gives our AI exclusive SKU-level insight into how people actually shop today.

 

Our proprietary DeepKNN model computes “embeddings” – vector representations in mathematical terms – that allow us to draw connections between shoppers, products and media signals. This allows models built on top of DeepKNN to surface intent patterns, and forecast likely buying outcomes based on real-world behaviours as opposed to educated guesses or broad demographics. We’re proud that DeepKNN received an industry award in 2024, a nod to both its approach and its impact.

 

The result of all of this is AI that goes beyond guesswork about behavioural trends and instead picks up the subtle nuances that shape each of our unique shopping journeys.

 

From predictive to agentic AI

 

Until now, AI has mostly played the role of advisor.

 

It crunched numbers, made predictions and helped guide decisions – but it stopped short of acting on its own. Agentic AI takes things a step further, with systems that can decide and act within guardrails designed by humans – specifically marketers in our case. 

 

Because the agentic world is so new, it might help to distinguish between the agentic and non-agentic elements of this change. 

 

The agentic aspect includes foundational capabilities that make the shift tangible, such as:

• Creating campaigns in minutes and running them: provide a simple natural-language brief and get a fully built, ready-to-run campaign – no more back-and-forth or manual setup
• Building smarter audiences with a prompt: describe who you want to reach, and AI automatically groups shoppers by intent and behaviour, continuously updating segments as those signals change
• Getting an instant understanding of how campaigns are performing: beyond basic metrics, marketers get clear, plain-English explanations grounded in actual products, categories and measurable commerce results
• The non-agentic aspect includes the tools and approaches that existed before the agentic era:
• Generating creative variations instantly: Formats, visuals and copy can be tailored to product feeds and performance signals in seconds
• Making changes in the moment: These adjustments have traditionally required human intervention, even though signals shift continuously

Agentic AI sits on top of these long-standing capabilities, co-ordinating and surfacing them rather than replacing them. The agentic layer not only allows us to think about new capabilities for marketers, but also to access and surface the AI tools developed over years at Criteo. The agentic paradigm – and particularly the MCP protocol – lets us envision how these tools will support marketers with even more intelligent assistance.

 

Retail media, and the wider commerce media landscape, is where this shift becomes especially visible.

The fast pace of retail environments, the depth of product-level signals and the need to optimise across thousands of SKUs make retail media a natural proving ground for agentic workflows. The ability to brief in natural language, generate optimised audiences on the fly and receive clear explanations of performance aligns directly with how retail media is planned, bought and measured today.

 

As agentic systems mature, commerce media will increasingly benefit from AI that doesn’t just analyse signals but acts on them – building campaigns, adapting strategies in the moment and ensuring spend follows real shopper intent.

 

Responsible by design

 

Adopting AI naturally raises some legitimate concerns: transparency, data privacy and the risks of “black box” decision-making.

 

Our AI Lab and its network of machine-learning experts – more than 120 specialists in total – run over 100,000 A/B tests each year to validate and refine our models. Every algorithmic decision is continuously checked to align with what marketers expect, and our models are refreshed as frequently as every six hours.

 

We also work closely with France’s INRIA (National Institute for Research in Digital Science and Technology) through FAIRPLAY, a joint research group focused on fair, privacy-preserving AI that protects users without dragging down performance. And because responsible AI can’t live in a vacuum, our cross-department Regulatory AI Group keeps a close eye on how emerging regulations shape the way we design our algorithms, even hosting a companywide Trustworthy AI symposium in 2025 to push the conversation forward.

 

These safeguards ensure that AI is not only effective, but also trustworthy.

 

The road ahead

 

There’s plenty of buzz around AI right now, but make no mistake – the shift toward agentic AI isn’t a passing trend. It marks the arrival of a new channel through which commerce media can be planned, bought and optimised, sitting alongside the ones marketers already use.

For brands, retailers and agencies, this opens the door to delivering more relevant experiences, reaching new audiences and driving stronger results – all while maintaining control and trust.  

---

For the complete picture of how the agentic transformation will unfold, read Powering AI – Assisted and Agentic Commerce: Criteo’s Vision for the Next AI Era

---

By Liva Ralaivola, VP Research, Head of Criteo AI Lab, Criteo