Milliseconds to Match: How Criteo Bids on a Billion Anonymous Profiles

Nathan Labenz opens The Cognitive Revolution with a constraint that has not changed in twenty years of Criteo’s business. From the moment a browser asks for a page, the company has milliseconds to locate the visitor among about a billion profiles, choose one product out of many millions, and turn that choice into a bid in a live auction. Win, and an ad appears. Lose, and the next request is already on its way. The system does this billions of times a day.

The people who built that loop are Diarmuid Gill, Criteo’s chief technology officer, and Liva Ralaivola, vice president of research and head of the AI lab. They are joined by Alex Persky-Stern, CEO of Waymark, who took over that company from Labenz and now partners with Criteo to supply AI-generated commercial creative. Labenz’s own view is that without affordable targeting, a lot of long-tail products would not be viable. Gill and Ralaivola stress relevance, free access to information, and how easy it is to opt out.

Labenz cues a Bernie Sanders sit-down with Claude, in which the senator asks what Americans should know about how their data is collected, profiled, and used. The tone is ominous. Gill says the industry still does a poor job of demystifying the stack, and then he draws a hard line: Criteo does not collect personal information. What it has is a random anonymous identifier, plus signals about products people have seen, liked, or disliked. The rest, he argues, is a value exchange. A system that knows nothing about you shows random inventory; the brain is excellent at filtering it out. A system that knows interests has a chance of being looked at.

Advertising, in his phrase, is the lubricant that keeps the internet open and free. It is the revenue that lets content and service providers stay off paywalls. Transparency and the right to leave are the conditions that make the bargain legitimate. Criteo was an early user of the Ad Choices icon: a click shows why an ad appeared and offers a one-tap opt-out. Once those two pieces are in place, Gill says, the arrangement can create value for the user, the publisher, and the advertiser.

Ralaivola gives the same job a more colorful name. The middle party, he says, tries to be like mentalists: very few cues, a hard inference about what is relevant right now, and a hope that everyone in the exchange is better off. As the intermediary, Criteo has the least data and the most difficult AI problem. That is exactly why he finds the work interesting. Labenz then asks the concrete question everyone has after the cookie banner: if you opened the file, what is actually in it, and what did those consent pop-ups change?

Gill walks the loop from the retailer site outward. When someone lands on a brand or shop using Criteo, the company can write a cookie with a random ID and no personally identifiable information. If that person looks at a phone and then leaves, the system remembers the interest. Later, on some other site, an impression comes up for auction. Criteo can bid the same phone, or it can treat the visitor as a tech enthusiast and show adjacent products. iPhone shoppers and Android shoppers, he notes, do not look the same as audiences.

Ralaivola reduces the decision to a valuation problem: should we bid on this placement, and if we do, how likely is a click and eventual revenue? The models are classifiers trained on the browsing, click, and purchase traces Gill described. More sophisticated deep models raise precision and bury explainability, a trade-off the whole industry is living with and that Criteo’s research group still treats as an open scientific problem. Utility went up; the ability to say, in plain language, why this ad won is what they now have to make up.

After a sponsor break, Labenz asks for the actual architecture. Gill says each opportunity is offered to several ad-tech delegates acting for brands. The bid is the expected value of a click-through that ends in a purchase. Inputs include products seen, purchase history, the context of the current site, the device, and on the order of 150 features in all. Models are trained offline; inference is cached and run at very low latency. The product to show is chosen only if the auction is won. All of it has to finish in milliseconds.

The worst week of the year is Black Friday and Cyber Monday. Gill asks listeners to imagine running a car at 300 miles an hour for a week, then returning it to ordinary traffic, on the same machinery, with no loss of precision. Criteo’s stack has to absorb about 300 percent of normal load and still bid correctly. Ralaivola’s summary of the dual constraint never changes: be accurate, be fast, and do it billions of times a day without falling over. His hat, Gill says, goes off to the engineers who keep that car on the road.

Persky-Stern asks whether chat agents, which already know a lot about a user, will change the profile itself. Ralaivola says the privacy questions are still open. What he will say is that conversational models are strong at general reasoning and can recommend shoes, but they do not hold Criteo’s commerce data. The technical problem is to merge the two: let the LLM do world knowledge and dialogue, and let Criteo’s models supply accurate product state. The interesting direction, for him, is not dumping chat history into Criteo so much as using those models to enhance the commerce models the company has spent years building.

Gill is blunt about why a frozen model fails in retail. An LLM is true at the moment it is trained. Commerce is not. Flash pricing around Black Friday and ruptures in stock move inside hours, while the large labs refresh knowledge by crawling the web and then watch it go stale. Criteo sits on a network of about 17,000 retailers and ingests product data daily, sometimes more than once a day. In a hybrid setup, a user who asks for a product gets the richness of an LLM and a price and availability that are still true. Clicking through to the wrong price or an empty shelf, he says, is a uniquely bad experience.

The upside is product discovery that finally resembles a good in-store associate: someone who knows the full catalog, will take follow-up questions, will name the good and the bad of each item, and does not get tired. Ralaivola adds that agent protocols such as MCP have made the plumbing easier. Criteo used to adapt separately to each surface; now the job is to expose twenty years of product tools in a protocol-compatible way so an LLM can call them. Gill later stresses that the OpenAI partnership is still early, that both companies are privacy-driven, and that they intend to send only the minimum context needed to show an ad, and only in contexts where an ad makes sense.

Ralaivola narrates the twenty-year encoding path. For a long time, visits, products, and clicks were packed into a huge sparse vector, somewhere between 2 to the 12th and 2 to the 20th dimensions, filled with zeros and ones, because sparse linear algebra was fast. On top of that sat logistic regression: a linear model, one neuron, hand-designed features. It scored auctions quickly. It did not survive cookie changes, new data shapes, or the arrival of language models. That unsustainability is why the Criteo AI lab was created in 2018, with a mandate to stop crafting features and start computing them from data.

The proprietary step is DeepKNN. It compresses those enormous sparse histories into roughly 200 to 1,000 deep-learning features, and other classifiers then train on that denser representation. The lab is already on the next hop, bringing in transformer-style models that adapt more readily as the data shift. Methods, loss functions, and model sizes are documented on the company blog and in papers, which Ralaivola invites listeners to download. Labenz compares the setup to Stripe’s payments foundation model: a costly encoding, amortized across many tasks, that developers can treat as a rich black-box input.

Criteo is on the same path, with three or four foundation models that embed products, user timelines, and related objects, then publish those vectors internally. Other teams use them as a warm start. A hackathon a month before the taping was the first time many groups got the embeddings; the reaction, Ralaivola says, was that the company already had gems and there was no reason to start from scratch. Versioning those embeddings like software, so a new release does not break every downstream model, is a large ongoing project. The last-click update still has to be applied online, which is where the engineering effort is heaviest.

Once products and users live in the same semantic space, several jobs collapse to nearest neighbors: similar products, audiences around a product, products around a user. Gill says those warm starts also matter commercially. Criteo made its name in lower-funnel retargeting and is now building mid-funnel products for customer acquisition and discovery across the open web, social, connected TV, and even LLMs. Anything that gets an advertiser to high return on ad spend faster, rather than from zero, is why the foundation models are exciting. His rule of thumb remains: given millisecond bidding, a fast good answer is worth more than a perfect slow one.

Gill, who is Irish and has lived in France for eleven and a half years, calls the claim that Europe cannot do AI “fundamentally wrong.” Criteo is headquartered in Paris, a fact Labenz suspects will surprise listeners who treat ad tech as an American industry. Being born under GDPR, Gill says, made careful handling of user data a first principle rather than a retrofit. The company operates globally, including a large U.S. market and offices in Ann Arbor, Toronto, and elsewhere, and it tries to go beyond local rules rather than run a looser stack outside Europe. The AI work, meanwhile, can hire from anywhere.

Ralaivola, a former university professor, points at France’s mathematics and computer-science training. Students there, he says, are strong at formalizing a problem, which is the skill you need to craft a model that answers a specific question. Mistral is one French name listeners already know. Gill, who studied engineering, adds the roll call of modern mathematics that underpins machine learning: Laplace, Lagrange, Fermat, Galois. That tradition, he argues, is a standing advantage Criteo has drawn on.

Labenz was startled to find about fifty faces on the AI lab’s public site. American labs, he jokes, fear a Zuckerberg blank check. Gill’s answer is culture and interesting problems: he commutes more than an hour to sit with people who are good at the work, and long tenure creates an ecosystem that is hard to poach wholesale. Competition keeps leadership honest. Ralaivola’s answer is professional: research scientists have to publish, be reproducible, and keep a public academic identity. Suppressing papers would make hiring the best people impossible. The lab also sponsors PhD students on real projects, some of whom convert to full-time roles.

On whose citizens are better served, Gill refuses a simple EU-versus-U.S. scoreboard. Criteo talks with data-protection offices on both sides and argues for transparency, consent, and opt-out, which is why it shipped cookie consent before the law required it. If targeting vanished, advertisers would stop paying, publishers would raise paywalls, and users would lose free content. After California’s CCPA and CPRA, he says, the two regimes have largely equalized and borrowed from each other. Criteo therefore runs one Euro-compliant approach worldwide, treating European birth as a portable design rather than a tax.

Creative has been the missing chapter. Persky-Stern’s experience at Waymark is that a lack of decent assets is one of the biggest barriers keeping smaller advertisers off Criteo, Meta, and Google. Gill agrees that generative models can both make ads more engaging and democratize production for the mid and long tail. Criteo recently launched a self-service product, Criteo Gold, aimed at those smaller advertisers, and he credits partners such as Waymark for making the campaigns look like something a brand would actually run.

Ralaivola is cooler on real-time generation. For years Criteo has done dynamic creative optimization: templates and visual assets assembled online like Lego. Generative image models are not yet fast enough to run in the auction; nobody will wait five seconds for a page. So assets are generated offline, by Criteo or by partners, and composed at request time. He guesses it may be two or three years before generation is quick enough, and that someday the last step might even run on a phone or a television. Until then, the split between offline generation and online assembly stays.

Asked how far personalization should go, Gill stays at the audience level. Brands still want control of look and feel, and he is not sure a unique ad for every person adds much utility for the shopper or the advertiser. The hyper-personalized cases that do make sense are user-initiated: you are invited to a wedding in the South of France and ask an agent for outfits and a virtual try-on, or you want to see a sofa in your own living room. “For advertising to be truly useful, it has to be trustworthy,” he says earlier. If it is creepy, it will not work. Ralaivola’s engineering version of the same rule is that you must not strip the user of the right to decide; every model review should ask where the human still sits.

On connected TV, Gill’s test is whether the ad is intrusive. Demos that composite brands into the scene of a sitcom or a match feel premature; pre-roll, mid-roll, and post-roll already have an accepted grammar. Cross-device identity is handled with the same caution. Logins on multiple devices can be treated as deterministic. Probabilistic matching is reserved for very broad targeting, and someone who has opted out should not be re-identified on another screen. Ralaivola’s longer dream is a small model in glasses or on a phone that applies a private last-mile tweak the cloud never sees.

Labenz puts two opposing bets on the table. If AI makes people richer, time gets more expensive and the market for a slice of attention should grow. At the same time, agents can evaluate every shoe in the catalog, which could collapse search costs and take bargaining power away from whoever was willing to bid on that time. Gill doubts a singularity is arriving on any useful schedule. What he does see is a better discovery loop: you state criteria to an LLM that is wired to live commerce data, you get a focused answer plus the accessories that go with it, and retailers reach people they used to find by guesswork.

Ralaivola pushes the thought further. Today people are exposed to ads whether they like it or not, even if they can opt out. With capable assistants, they may start requesting advertising as a filter. You tell the agent you want shoes, ask it to look at a hundred pairs, and have it put six in front of you. A new intermediary appears in the stack, and “show me five trips to Greece, no more, no less” becomes a normal command. In five years, he says, it is plausible that someone will simply ask to be shown ads. Persky-Stern’s line is the one Labenz ends on: if the advertising is good enough, it might be better than search, and the border between the two is already collapsing.