Fresh Product Data Is the Constraint for LLM Commerce Discovery

Criteo’s core advertising system is built around a narrow, repeated decision: when a person loads a page and an ad opportunity appears, should Criteo bid for that placement on behalf of an advertiser, what should it bid, and which product should appear if it wins?

Diarmuid Gill describes the data foundation as intentionally limited. Criteo does not collect personal information, he says, but creates a random anonymous ID, usually through a cookie when a user visits a retailer or brand site. That ID can then be associated with commerce signals: products viewed, products clicked, products purchased or not purchased, and broader interest categories inferred from those actions. A user who looks at a mobile phone might later be treated as part of a tech-enthusiast audience; the type of phone can also matter, because it suggests different affinities.

The operational setting is real-time bidding. A publisher has an available impression. Multiple ad tech providers receive the opportunity. Each acts as a delegate for advertisers and must decide whether that particular user, on that particular surface, in that particular context, is valuable enough to bid on. Gill reduces the commercial logic to expected value: how likely is this person to click, return to the advertiser’s site, and ultimately buy?

Liva Ralaivola puts the machine-learning task in its simplest form as a binary classifier: should Criteo bid on this placement or not? The real system is more complex, but the abstraction matters. The models are trained to estimate whether a placement, paired with a product or advertiser, is likely to create revenue. They are not producing long-form language or reasoning step by step like a conversational model. They are scoring opportunities at extreme speed.

Gill says the model may take in roughly 150 features, including the products a user has viewed, prior purchase history, the context of the website, the device, and other signals. These are fed into trained models that estimate click and purchase likelihood. If Criteo wins the auction, the system must also choose what product to show and how to render the creative.

All of this happens in milliseconds. The models are trained offline, but inference must happen in real time, under latency constraints and at very large scale. Gill adds another constraint: the same system must handle seasonal spikes such as Black Friday and Cyber Monday, when traffic can run at multiples of ordinary volume.

His analogy is a car that must run for a week at 300 miles an hour and then return to ordinary operation, using the same machinery. The engineering challenge is not only model quality. It is quality under latency, reliability, and peak-load pressure.

Liva Ralaivola frames the evolution of Criteo’s models as a shift in representation. The old problem was not only how to classify an ad opportunity, but how to encode the raw facts of browsing, products, and clicks so that a model could use them.

Earlier systems used very large sparse vectors. Ralaivola mentions encodings with sizes such as 2^12 or 2^20, containing many zeros and some ones. Sparse representations made fast computation possible, and Criteo could train logistic regression models on them. In deep-learning terms, he describes this as analogous to one neuron with many inputs: a linear model over a very large hand-built representation.

That approach worked, but it depended on manual feature crafting. Engineers and data scientists had to decide how to represent products, sites, and user histories. Ralaivola says one reason Criteo created its AI Lab in 2018 was to move beyond this pattern. The goal was to bring deep learning into the system so features could be learned from data rather than hand-designed every time the environment changed.

The current architecture uses learned dense representations. Ralaivola refers to Criteo’s proprietary Deep KNN methodology, which computes deep-learning features from the underlying commerce data. Instead of sparse vectors with thousands or millions of dimensions depending on the encoding scheme, the system can work with hundreds to around a thousand learned features. These representations then feed other models that perform classification and recommendation tasks.

The tradeoff is familiar in modern AI: better utility can reduce explainability. The more sophisticated the model, the harder it becomes to explain exactly what it computed. Criteo, like the broader industry, has gained accuracy and relevance through deep learning, but must make up ground in explainability. Ralaivola identifies explainable AI as an active research topic, including within Criteo.

That tension also shapes user-facing transparency. A simple AdChoices explanation may tell someone they saw an ad because they are interested in skincare or technology. But the actual model may be combining many signals through learned representations that do not map neatly to a single human-readable reason. Ralaivola’s point is not that transparency is impossible, but that high-utility deep models and high-readability explanations are difficult to optimize simultaneously.

Criteo’s foundation-model work is not presented as a single universal model that directly solves every advertising task. Liva Ralaivola says the company is building several foundation models — “three or four,” in his description — that compute embeddings for different commerce objects, such as products, user timelines, and related entities. Those embeddings are then made available across the company.

The architecture resembles a modular internal platform. A team working on a new model does not have to start from raw data or handcrafted features. It can begin with vectors that already encode a large amount of commerce signal. Ralaivola says teams used these foundation-model embeddings during a recent internal hackathon as a “hot start” or “warm start” for new models.

Diarmuid Gill connects this to Criteo’s product expansion. Criteo’s original strength was retargeting, which he describes as lower-funnel advertising. The company is expanding into mid-funnel use cases such as customer acquisition, product discovery, multichannel campaigns, open web, social, connected TV, and LLM surfaces. In that context, the value of shared embeddings is speed to performance. A new product or model can start closer to effective return on ad spend rather than learning everything from zero.

Ralaivola gives a concrete account of how embeddings support recommendation. If a product is embedded into a vector space, and similar products land nearby in that space, then recommending related products becomes a similarity search problem. The same applies to users. If user representations and product representations are semantically meaningful, Criteo can find users close to a product and build an audience, or find products close to a user and recommend them.

That does not remove the latency problem. Nathan Labenz presses on what must be precomputed and what must happen at runtime. Ralaivola agrees that much can be cached: base user profiles, product encodings, and other embeddings. But the most recent user action may be critical. Incorporating that fresh signal requires online adjustment, sometimes using faster versions of larger models, while preserving the geometry that makes similarity comparisons useful.

Gill describes this as a constant balancing problem. With unlimited time, a system could produce a better answer. In real-time bidding, it must produce the best possible answer within milliseconds.

This is why Criteo’s AI challenge differs from the popular image of foundation models. The problem is not only training a powerful model; it is making the model’s representations operational inside an auction system that must respond billions of times a day.

Criteo’s OpenAI partnership is still early, and Diarmuid Gill and Liva Ralaivola are careful not to over-specify how it will work. What they do describe is the prospective division of strengths between conversational AI and Criteo’s commerce infrastructure.

LLMs, Ralaivola says, are strong at general reasoning and can already perform a kind of recommendation. A user can ask for shoes, and a model can suggest plausible products. But Criteo’s contribution is commerce data: what products exist, what people are interested in, and what commercial conditions are current. The technical challenge, as he frames it, is combining the LLM layer with Criteo’s models so that general reasoning can be enhanced by accurate commerce information.

Gill stresses freshness. An LLM trained on crawled web data may be accurate at the time of training or updating, but product data changes quickly. Prices shift, especially around events such as Black Friday. Inventory runs out. A recommendation that points to an out-of-stock item or a stale price creates a bad user experience.

Criteo’s network gives it a different signal. Gill says the company works with 17,000 retailers and ingests product data daily, sometimes multiple times per day. In the hybrid architecture he expects, the LLM contributes language, reasoning, and conversational product discovery, while technology from a company like Criteo can help keep product information current and accurate.

Gill’s preferred analogy is not a search results page, but a strong in-store sales assistant. A good assistant knows the catalog, answers follow-up questions, explains tradeoffs, and leaves the customer feeling understood. LLMs paired with accurate product information could provide that experience online: a user can ask, refine, compare, and drill down without the system becoming tired or stale.

Ralaivola adds a deployment point, while keeping it at the level of a developing integration pattern rather than a settled product architecture. Agentic protocols such as MCPs, as he describes them, make it easier to connect existing tools to LLMs. Instead of adapting separately to each surface or website, Criteo’s task is to make its tools available in ways that can work with those emerging protocols. The technical detail matters because it may lower the integration cost for commerce systems to appear inside conversational interfaces.

The partnership also raises questions about what signals are exchanged, and Ralaivola says Criteo is still investigating the right way and the right data to exchange. He describes scenarios the AI Lab has explored: using full conversation text, using summaries of conversations, or using vector representations of conversational context as an additional signal. He gives an internal example from earlier work: a Criteo employee built a Slack-based conversational agent using messages about account strategies to answer troubleshooting questions for clients. It was not product recommendation, but Ralaivola sees the structure as similar: infer a need from conversation and recommend a relevant solution.

Gill is cautious about the operational form. The OpenAI partnership is early, and both companies, he says, are privacy-driven and focused on user consent. Criteo should receive only the information needed to show an ad at a given time, and not every conversational context should lead to advertising.

Diarmuid Gill repeatedly returns to a value-exchange model of advertising. Personalized advertising is useful, he argues, when it gives users more relevant experiences, gives advertisers efficient access to likely customers, and funds free or open access to content and services. If personalization disappears, advertisers spend less because their products are shown to people with no interest. Publishers and service providers then need other monetization models, including paywalls.

For Gill, that value exchange only works if users have transparency and control: to understand why an ad appeared and to opt out. He points to Criteo’s role in the AdChoices icon, which lets users click through, see why they saw an ad, and opt out of personalization.

Liva Ralaivola adds that Criteo often has less information than other parties in the chain and a hard prediction problem. He describes Criteo as a “middle party” with few cues, trying to infer relevance from limited signals. That makes the AI problem more interesting to him, but it also bounds the data relationship: Criteo is not claiming omniscience about users.

The privacy discussion becomes sharper when conversational agents enter the picture. Ralaivola says people may share massive amounts of information with assistants, which makes trustworthy machine learning more practical and urgent. His examples of trustworthiness include fighting hallucination, preserving privacy, choosing not to answer in some contexts, and recognizing when a user should not receive recommendations. He presents this as both research and preparation for future regulation or expectations.

Gill’s practical rule is blunt: if advertising feels creepy, it will not work. It has to be useful, trusted, and engaging. Creepiness undermines the commercial objective because users will not respond positively.

Cross-device identity shows the same boundary. Labenz asks whether matching a person across phone, desktop, shared Wi-Fi, mobile networks, and household devices has become an AI problem. Gill answers first with consent. Deterministic signals, such as being logged in across devices, can establish the same person. Probabilistic signals may be used in more limited, untargeted ways. But if someone has opted out of advertising, Criteo does not want to re-identify and target them through another route.

The company’s European origin matters here. Gill says being born in Europe, under GDPR-era sensitivities, made privacy a first principle. He rejects the idea that “you can’t do AI in Europe.” On regulation, he says Criteo tries to build global solutions, and that what it has done in Europe is used globally rather than treating other markets as places to be looser. He also notes that California rules such as CCPA and CPRA have narrowed some differences between US and EU regimes, and that Criteo engages with data protection authorities in both regions.

Nathan Labenz raises a common perception: ad tech is often thought of as American, and Europe is often portrayed as a difficult environment for AI. Diarmuid Gill and Liva Ralaivola reject both implications.

Gill, an Irish executive who moved to France, says Criteo benefits from the quality of European AI and data-science talent. Ralaivola, a former professor, is more specific about France. He points to a strong mathematical and computer-science tradition, and argues that formal training matters because AI requires the ability to formalize, model, and select the right technical tools. He says France has students and engineers trained in exactly those foundations.

Gill extends the point historically, naming French figures such as Laplace, Lavoisier, Fermat, and Galois as part of the mathematical tradition underpinning modern machine learning and AI. The claim is less that history itself creates models, and more that France has a culture of formal mathematics from which companies like Criteo benefit.

Criteo’s AI Lab is also intentionally visible. Labenz notes that the lab’s website lists roughly 50 people, which surprised him because US AI companies often avoid public researcher rosters for fear of poaching. Gill’s answer is culture and problem quality: people stay when they have difficult, meaningful problems and strong peers. Competition for talent forces leadership to keep Criteo a compelling place to work.

Ralaivola adds a research-science norm. Research scientists publish. They maintain websites, show their work, and make research reproducible. Criteo wants its researchers to operate in continuity with academic and industrial science, not disappear into a closed corporate system. Publication helps researchers’ careers and helps Criteo remain connected to the broader AI ecosystem.

The company also sponsors PhD students, Gill says. Those students work with Criteo researchers on real-world projects, publish their research, and sometimes become full-time employees. Not all do, which Gill presents as acceptable; the broader point is sustaining a pipeline between academic work and deployed systems.

That openness is tied to Criteo’s view of its moat. Labenz compares Criteo to Stripe’s payments foundation-model work: companies with privileged data networks can often discuss methods publicly because the data flow and operating context are difficult to replicate. Ralaivola says building foundation models is hard, and Criteo has the data. The implicit position is that publishing techniques does not give away the full system.

Creative enters the discussion as both a bottleneck and a new surface for AI. Nathan Labenz and Alex Persky-Stern raise the problem from the advertiser side: many businesses cannot produce enough high-quality creative to participate effectively in ad platforms. If generative AI lowers the cost of creative, it could expand the advertising market, especially for smaller advertisers.

Diarmuid Gill sees generative AI as one of the most exciting parts of the advertising stack because it can make ads more engaging and democratize content creation. Historically, mid- and long-tail advertisers could be excluded by the cost and complexity of producing high-quality assets. Gill says Criteo’s self-service product, Criteo GO, is meant to make it easier for smaller advertisers to create campaigns with strong creative. He identifies Waymark as one partner in that effort.

Liva Ralaivola separates two problems: generating assets and assembling ads in real time. Criteo has long used dynamic creative optimization, where templates and visual assets are arranged at runtime. He compares the assets to Legos. Generative AI can help create those Legos offline, either by Criteo or partners, but current image generation is not fast enough to create a fresh ad from scratch during a page load. Waiting several seconds for a page to load is not compatible with ad-serving constraints.

So the near-term system is hybrid: generate or prepare visual assets offline, then assemble and render them online using Criteo’s existing dynamic creative infrastructure. Ralaivola speculates that, in the future, generation may become fast and distributed enough to happen on a TV, phone, or other local device, but says that is not where the technology is today.

Personalization has limits even when the infrastructure can support more. Stern asks whether advertising will become individually personalized, with a distinct story for each person, or remain at audience and context levels. Gill expects most useful personalization to remain audience-level. Brands still want control over look, feel, and representation. More importantly, he is not convinced that person-by-person creative variation creates enough utility for users or advertisers.

Ralaivola offers a different long-term possibility: shared computation. Criteo might provide an audience-level ad, while a personal device performs the final adaptation locally, using information that never leaves the user’s device. In that version, the most intimate personalization happens on the user side, not Criteo’s.

Gill sees the clearest case for hyper-personalization when the user initiates it. If someone asks an assistant for outfits for a wedding in the south of France, or wants to see how furniture would look in their home, the personalization is not creepy because the user asked for it. Virtual try-on and home visualization become advertising-adjacent experiences because they provide utility rather than interrupting attention.

Creative contextualization also has boundaries. Stern asks about connected TV and matching ad creative to the tone, genre, or even specific moment of a show or movie. Gill’s guiding principle is non-intrusion. Advertising should feel relevant and seamless, not interfere with the content. Pre-roll, mid-roll, and post-roll are familiar formats; inserting ads into content itself may be possible, but Gill is cautious about whether users are ready for it.

Liva Ralaivola’s broadest principle for AI systems is “human in the loop.” He applies it not only to creative generation, but to bidding models, agentic systems, and AI projects generally. The challenge is not simply keeping a human somewhere in the process. It is placing the human at the right point.

If automation strips the user of the right to decide, Ralaivola says, the system creates problems of liability, responsibility, and acceptance. But if the system demands human decisions at every point, it fails to deliver the benefits of automation. The design problem is to determine where human judgment belongs so the model can operate effectively without displacing agency.

This principle also clarifies Criteo’s posture on generative creative. Ralaivola says Criteo does not currently build its own creative models, partly because they are costly and expensive. Teams are working on integrating creative capabilities over multi-year horizons, but quality control, brand guidelines, and advertiser intent remain central. Diarmuid Gill says advertisers and retailers have their own look and feel, and Criteo’s goal is advertising that works within those constraints.

The “human in the loop” theme also connects back to conversational agents. Ralaivola thinks conversational agents work well partly because they heavily rely on human interaction. The user asks, refines, chooses, and rejects. That interaction is not a weakness of the system; it is part of why the system is useful and legitimate.

Nathan Labenz suggests two opposing forces for advertising as AI becomes more capable. If AI makes people richer and their time more valuable, competition for attention may intensify, benefiting advertising. But if AI agents dramatically reduce search and matching costs, purchasing decisions may depend less on who bid for attention and more on how thoroughly a user’s agent evaluated the market.

Diarmuid Gill focuses on better product discovery. A user with specific criteria could ask an LLM-like interface for the best product, regardless of price or with any other constraints, and the system could combine broad reasoning with rich commerce data. In that world, advertising helps retailers surface relevant products in contexts they could not previously reach, while users get more complete information for decisions.

Liva Ralaivola imagines a deeper change in the value exchange. Today, users are often exposed to advertising unless they opt out or avoid the surface. In an agent-mediated environment, users may actively request exposure. They might tell an assistant: look at 100 pairs of shoes, filter them down to six, and show me those six. Or: show me five trips, no more and no less. Advertising becomes part of controlled discovery.

Labenz frames this as a possible collapse of search and advertising: if advertising is good enough, it may become better than search. Ralaivola agrees with that framing, but the point remains speculative. The intermediary might change. The user’s assistant may become the actor that receives, filters, and ranks commercial options before the user ever sees them.

That possible future also reinforces the earlier constraints. The assistant needs fresh product data, not stale catalog knowledge. The user needs control over how much they want to see. The system has to avoid creepiness by making personalization responsive to user intent. And Criteo’s role, if Gill and Ralaivola’s view proves right, is not merely to place ads into existing surfaces, but to supply the commerce intelligence that lets agents discover what is available, relevant, and purchasable now.