--- "title": "Why the future of commerce AI isn’t one model" "date": "2026-10-15T12:25:13-04:00" "url": "https://www.criteo.com/blog/why-the-future-of-commerce-ai-isnt-one-model/" "description": "Discover how Criteo applies specialized models to improve bidding, product recommendations, and decision-making across the shopper journey." "image": "https://www.criteo.com/wp-content/uploads/2026/10/Blog-Post-Header-Image-DESIGN-v3.png" "author": "criteo_multi_admin" "category": "Commerce AI" --- At any given moment, hundreds of AI models are running inside Criteo’s performance engine.   Some predict whether a shopper is likely to buy, while others decide which products to show, where to set the bid, or which creative to choose. Each one is built for a specific job and a specific kind of dataset, and each one is continuously retraining as shopping behavior changes.  Recently, two new deep learning systems went into production in that engine: one that changes how Criteo bids, and one that changes how it chooses which products to recommend. Together they deliver more than 2% uplift in advertiser value at constant spend.  These aren’t Criteo’s first deep learning systems. DeepKNN, which pairs deep learning with the ability to search by meaning instead of keywords, has powered most of Criteo’s retargeting recommendations since 2023, and a multi-task deep learning model rolled out in November 2025 delivered a 3% ROAS uplift for participating advertisers.   The more useful question about any commerce AI, though, is how it decides which technique to use where. Criteo runs several kinds of AI side by side, from deep learning to simpler, faster models, and uses whichever is best suited to the decision at hand. Choosing well, one decision at a time, is most of the work.  ### **Why does commerce AI need more than one model?**  Every ad opportunity happens in an instant, but each one depends on a chain of separate decisions:  - Value: Is this impression likely to lead to a sale, and how much should the system bid?  - Relevance: Which product is most likely to match what this shopper wants right now?  - Efficiency: How should budget move across campaigns, channels, and audiences to improve results?  - Measurement: Did the campaign drive real, incremental impact?  These decisions are connected, but they run on very different kinds of data. Predicting what an impression is worth means weighing hundreds of structured signals in a few milliseconds. Understanding what a shopper wants means reading a sequence of behavior over days or weeks and working out where it’s heading. A model built to do the first well is not the one you would build for the second, and a platform that applies one architecture everywhere is making some of those decisions worse than it needs to.  That is why great commerce AI depends on specialization. The best systems start with the decision that needs to be improved, then apply the model best suited to that job.  ### **Two new models for two different decisions**  Both of Criteo’s new models use deep learning, applied to two very different commerce decisions. **1. Gandalf: A deep learning model for bidding** Every bid is a prediction about value. Gandalf is a new deep learning model built to handle the dense data that bidding requires, evaluating hundreds of signals about the impression, context, shopper, and advertiser, all in milliseconds.   Criteo still uses simpler models for plenty of decisions, but bidding is one where deep learning now does the job more effectively.   The result is a more accurate estimate of what each opportunity is worth, helping the system determine when to bid up, when to prioritize efficiency, and how bidding decisions improve over time. In practice, that means paying less for impressions unlikely to become sales, and more for those that will deliver outcomes. **2. Transformer recommendations: A deep learning model for product selection** Product selection is a different challenge, as a shopper’s interest is rarely captured in a single click. People browse, compare, leave, come back, and narrow down before buying.  The new transformer-based recommendation model can read those shopping actions as a sequence, learning a shopper’s journey the way a language model learns a sentence, where each action takes its meaning from what came before it. That learning enables the system to predict which product a shopper is likely to want next, not just the one they already viewed.  Apart from both of these models leveraging deep learning, they share very little: different data, different architectures, and a different definition of a good answer. Both are highly specialized to a specific task, pointing to the value in employing different models for specific jobs.   ### **What happens when the models work together?**  Gandalf estimates what an impression is worth, while the transformer model decides what to show in it. Because those decisions happen in sequence, improving one improves the next: a sharper value estimate puts budget behind the right impressions, and a better recommendation makes each of those impressions more likely to convert.  The power of this approach is cumulative. Each model improves a specific decision, and because those decisions feed each other, the improvements stack across the shopper journey.  “*People often talk about AI as if there’s one technology that drives performance. In reality, performance comes from combining different types of models and applying each one where it creates the most value. At Criteo, our focus has always been building systems that make better decisions across the shopper journey, not optimizing for a single AI technique.”* — Nicolas Chrysanthos, Engineering Director, Criteo  Criteo has been using AI to drive performance outcomes for more than fifteen years, and the Criteo AI Lab has published research at AI conferences like NeurIPS, ICML, and RecSys since 2018. While technology and techniques have evolved several times over during that period, our thesis has been constant: start at the decision, choose the model that improves it most, and connect it to everything else.  **Ready to see how specialized AI models can work together as a system to improve performance?**[Talk to a Criteo expert.](https://www.criteo.com/talk-to-an-expert/)