With commerce data in easy reach and a strong programmatic foundation, Criteo is swinging hard at the AI opportunity in advertising.
CEO Michael Komasinski, who joined the company in February 2025 after leadership roles at Dentsu, Merkle and Nielsen, is focused on unlocking the power of the company’s intellectual property including its well-developed recommendation technology.
His company’s product roadmap is multi-threaded and includes bets on agentic commerce (Komasinski prefers AI-assisted commerce), self-serve advertising and a high-profile partnership with OpenAI for its ChatGPT ads product.
Criteo is ready for its AI moment.
Mr. Komasinski spoke to tipsheet about Criteo yesterday, including:
- Criteo and the ChatGPT ads pilot
- Conversational ads and performance
- Audience-based targeting in ChatGPT
- The RecSys opportunity vs. LLMs
- Defining—and redefining—agentic commerce
- Criteo’s agentic commerce strategy
- Agentic commerce recommendation services
- Retail media’s opportunity ahead
- AI shopping assistants as a product
- Self-serve and Criteo’s GO platform
- Performance advertising expansion plans
- The future of the agency model and SaaS
- Criteo’s long-term outlook
Scroll down for the interview, which has been lightly edited for clarity.
tipsheet: What’s the latest with Criteo’s participation in OpenAI’s ChatGPT ads pilot?
Michael Komasinski: We remain excited to be their first ad tech integration partner. And what that means is: we are integrating their ChatGPT inventory into our existing performance platform and treating that as another source of inventory in our cross-channel performance setup.
The response from clients has been outstanding — both existing and new — and we continue to onboard new clients to ChatGPT’s ad platform.
The way that works is pretty simple. We’ve got an API connection with ChatGPT — and there are aspects of onboarding that we handle, and then hand off to them to handle. The relevancy engine, or ad serving logic, sits with ChatGPT.
We’re working on ways to measure clickthrough and performance of ads as they occur so that our clients can understand how it fits in their overall performance makeup. We’re really encouraged with the traction that it’s getting.
And last week, OpenAI announced that they’re going to start to expand ads internationally, and we’re a great partner for that — we’ve got a global footprint and we’re going to continue to drive that partnership on a broader scale.
Can you talk more about performance of ads within ChatGPT?
We put out a piece of thought leadership recently, and one of the things that we referenced in there is that traffic coming from ChatGPT and similar platforms converts at a 1.5 times higher rate than typical traffic. And so that reinforces the point that you’re getting very highly qualified, high intent traffic from these ads.
What about the performance of advertising within the conversation of ChatGPT?
We don’t have anything to share at this time. I have read things in the press this week where some folks did studies on when ads were being served in a conversation. Were they being served in the first or second prompt? — or were they being served as late as the 13th or 14th prompt? I, like most people, found that super interesting.
What the insight introduces more broadly is the topic of conversational ads which happens outside of ChatGPT. This is a format that we’re piloting with some of our retailers.
Conversational ads introduce a Z-axis, if you will. If you think about ads in a two-dimensional context of targeting and creative, there’s now a Z-axis conversationally… such as “Why am I being shown that item?” or “Let me see that in blue” or “Yes, that looks cool. What price point is that?”
It introduces a really interesting third axis in how people interact with ads, how you think about monetization opportunities and how you think about ad serving.
There’s a lot there.
What can Criteo bring to the targeting of ChatGPT ads? Can it bring audience-based targeting, for example?
Not currently. That really depends on how OpenAI evolves the platform.
If they evolve the capability to receive that information, then we certainly would be ready to do that. But that’s not how it’s set up today.
Speaking broadly about conversational ads, Criteo has seen an opportunity with what’s known as RecSys (Recommendation System) versus the semantic world of large language models (LLMs). Can you speak to that opportunity and what it means for Criteo?
At its core, even the conversation demonstrates how different Criteo is as an ad tech platform. What I mean is… a more traditional demand-side platform (DSP) or sell-side platform (SSP) would be more of a workflow tool for building audiences, finding optimal supply paths and running campaigns.
The backbone of Criteo is about what’s the shopper graph, what’s the product graph and what are the impressions that are most likely to create the desired outcome for a campaign. There’s a highly-scaled recommendation engine that powers that. It’s really a prediction engine – essentially what are the optimal impressions to acquire to create the outcome?
And in order to assess the optimal impressions, especially cross-channel, you need high volume tests that build a graph and give you that intelligence to predict with accuracy.
I think that’s really interesting in the context of algorithms today because large language models, by definition, are semantic models. They’re great at predicting the next word in a sentence, how to string sentences together to build concepts and so on. Then, the information that they retrieve to do that happens through RAG (retrieval-augmented generation), which is about crawling sources — mostly on the open web, but any other dataset that they have access to — and then assembling it using a semantic learning capability.
A RecSys engine is built on data. So the simpler way to think about all this would be a RecSys engine is much more like behavioral research as opposed to attitudinal research. So not what people say, but what they actually do and at a very large scale – like Criteo.
We’re seeing a trillion opportunities to bid on every day. We see $3 billion of commerce transaction volume per day. And between the trillion impression opportunities and the $3 billion of outcomes, you can start to build sophisticated intelligence of what someone was shown, what was the creative, the time of day, the platform and what they did or did not do.
When you start to build that over billions and billions of opportunities, you create a highly tuned prediction engine that gets you even better results. And it tends to compound over time.
This is the same type of engine that Meta has built for ads, Netflix with movies, the engine that powers the interface of Airbnb and platforms like Spotify. These are the RecSys backbones that power, in all those cases, best-in-class user experiences. In the case of Criteo — highly performative, highly targeted advertising.
Let’s turn to agentic commerce, you’ve publicly agreed with Mobile Dev Memo analyst Eric Seufert that you believe “agentic commerce is a mirage.” Before we get into that, how do you define agentic commerce?
I’ve been thinking about this phrase “agentic commerce” in the last week or two. I think we need a vocabulary change in the industry, but it’s not easy to figure out what it should be.
It really should be called “assisted commerce.”
Unfortunately, “assisted” doesn’t anchor the AI component of “agentic commerce” which is why this debate has been manifested.
“Agentic” is derived from the word “agency.” And agency is a word that denotes a level of autonomy or acting on someone’s behalf — but that gets you down the wrong path.
I don’t believe in agentic commerce. I believe strongly in AI-assisted commerce, which doesn’t roll off the tongue as an easy phrase.
So, I think this comes down to semantics because there are some people that really do believe in these autonomous use cases. I’m not one of them.
People who do believe it completely underestimate human behavior and that shopping is this chore to be outsourced or eliminated at every turn — I disagree with that. Shopping is an expression of identity and an enjoyable experience where people are able to spend the money that they earn from working during the week.
Also, it’s not in retailers’ or other commerce players’ best interests to have a bot-to-bot transaction off a retailer’s site. The ability to serve customers and create intimacy and loyalty is surrendered to another party.
The autonomous use case is never happening.
How is Criteo addressing agentic commerce, or AI-assisted commerce, in the marketplace today?
One way has been by being an early mover and partnering with industry leaders like OpenAI. We’re also building capabilities and partnerships around the part of this new behavior that we think will thrive — and then avoiding the parts we don’t.
What will thrive, clearly, is this expansion and fragmentation of discovery. This is a big theme for us: discovery and its total addressable market.
Discovery is growing rapidly by the fact it encompasses traditional search and “agentic” — it’s a bigger TAM almost by definition.
Consider Google’s Gemini, traditional search, ChatGPT and then all the variants of search and discovery that have been happening in the last couple of years such as Reddit for peer-level information, Amazon and other retail media networks for Product Search and so on. AI platforms have accelerated it — so a bigger, more fragmented TAM. And for the ad tech industry, there are more opportunities to partner and “play” in that ecosystem. It’s good for companies like Criteo.
For our roadmap, we want to operate full funnel and cross-channel. And now that the discovery layer is available and growing, we want to connect that to driving outcomes in the lower funnel. In order to do it effectively, you need to be able to do that across channels because people are not channel specific. It significantly shapes our roadmap and strategy.
Can you talk about how Criteo’s new agentic commerce recommendation service fits?
Yes — the recommendation service that we launched has a couple of different applications.
First, we wanted to expose that capability in a very partner-friendly way in order to demonstrate the IP that sits beneath the business. We are testing with a couple of large platforms. And even if those partnerships don’t come to anything, it powers other use cases for the business.
So the way that we power relevancy decisions for retail media — i.e. when to serve an organic ad versus when to serve a sponsored ad — a lot of that depends on the density of the demand and the variation of the products themselves relative to the query that’s driving that opportunity.
It’s the same sort of intelligence that decides, “OK, this is a low density auction situation. We should serve organic only to preserve user experience and associated gross margin of closing that sale.” Or “this is a very dense auction with lots of homogenous products. We should allow sponsorship in here because we’re going to get a sale no matter what. And we have the opportunity to monetize around that in a way that adds value to a shopper.” It’s the same type of intelligence which powers that decision. And obviously that would be core to our retail media engine. We call that page intelligence.
Drilling in on the recommendation capabilities and its commercialization, do you consider that a business opportunity in itself — to recommend? Or is it more about spreading that core technology across other opportunities?
It’s the latter. And it provides the chance to deepen partnerships with industry leading AI platforms, which is great for Criteo.
The fact that we’re testing that with a couple of large platforms is testimony to the capability itself. If that results in a revenue stream, that’s great. If it just deepens the partnership and leads to other opportunities, well that’s great, too. Time will tell how that plays out. In any event, it’s our recommendation or prediction engine as we discussed and it’s the foundation of the company. Exposing it and opening up those conversations is really valuable for us.
What’s the agentic commerce opportunity that you’re discussing with your retail media clients today?
It starts with pushing back on the bear case that agentic commerce — or AI-assisted commerce — decreases opportunities for retail media monetization.
That is just patently untrue. It may shift, but the opportunities for retail media monetization are as great or greater in this new paradigm than they were in the past.
First, retail media networks represent aggregated demand. And advertising in a platform like ChatGPT represents a form of off-site retail media that today would happen with off-site display or programmatic. But now it can be extended into that channel as well. That’s really seamless.
Next, as we discussed earlier, if you support a shopping assistant on-site, you will make the same decisions about what to show in that prompt engagement user experience, very similar to how you decide whether to show a sponsored ad next to an organic ad on a product listing page.
This is not dissimilar to how ChatGPT is making a decision about when to serve an ad. There’s a relevancy and decision engine that decides when’s the right opportunity to serve a sponsored ad or an organic ad, and retailers now will have that opportunity as they set up their own shopping support in their native environments.
Also, with conversational ads — I create interaction either through a prompt or through an off-site ad that’s conversational. I’ve got the same relevancy engine running behind it to say:
- What is the feedback or the prompt that I’m getting from a user?
- And then, how do I give them the best experience, the best recommendation?
- And do I have an opportunity to bring monetization into that where the objectives overlap in a suitable way?
The other thing that will happen with retail is if the discovery layer becomes a big traffic driver, people will start to land in different places than they have before.
One of the points we’ve called out recently is that the product page is the new home page. As a retailer, that’s great. Now you’ll do different things with that landing page. You might enable more conquesting or invest more in display monetization around that landing event. You might focus more on basket building.
There is a multitude of opportunity to both drive commerce outcomes and media outcomes at the same time, even with that landing page because the idea that people are coming in “hot” and essentially concluding a single transaction with no interaction with advertising, or not paying attention to the retail shopping environment around them, it’s just false.
Thinking as consumers, if we land somewhere, we’re curious about what else is there. I would argue that those opportunities are even greater because you’re serving higher qualified, higher intent traffic.
Can you see AI shopping assistants becoming a product for Criteo?
Not in and of themselves, but certainly the intelligence that powers them. It’s already happening.
We’re piloting that with several retailers already, and like I said, it’s just an extension of the ad serving and page intelligence capability that we have today but served in a new environment. There might be a tighter carousel or fewer slots, but it’s the same determination of: what is this person asking for, what’s the optimal organic result and is there density sufficient to provide sponsorship alongside that.
Moving to Criteo’s GO platform, you launched general availability of a self-serve version of GO in the US and UK earlier this week. Why is this the right time for the self-serve model for GO and Criteo?
There’s a broad theme running through the advertising sector around the democratization of SMB advertising. Generative AI tools were probably the first unlock around that vector. But now it has spread to workflow, audience building, targeting optimization and the entire workflow of advertising. We want to play a part in that.
Criteo traditionally has had more of a managed service footprint. The self-service product potentially makes the product portfolio more scalable and allows us to go after a segment that was difficult to scale in a managed service context. So for many reasons, it’s a great time to launch the self-serve product.
What’s differentiated about GO is the cross-channel setup and its ability to deliver constant campaign performance. It works across display, native, video and social, currently. And we plan to add additional channels with consistent measurement rolled up to a campaign objective and that is differentiated in the market.
How is the performance side of Criteo’s business evolving? Does it start with GO?
We always try to simplify the strategy for the performance segment into three simple tentpoles: full funnel, cross-channel and self-service.
We’ve talked about self-service but add that “self-service agenda” also applies to other products.
We’re generally deprecating most of our user interfaces and moving everything towards API-first. Criteo as a platform is not heavily invested in user interface and that is an effect of AI platforms. Everybody wants to transact conversationally and you use the agentic front-end for that.
Using MCP as the connector from the front to the back, we’re re-engineering the product portfolio around that paradigm. It’s not a heavy lift because we were always back-end heavy, front-end light.
So, GO is the flagship, but it runs deeper than GO in the company.
In terms of the full funnel, cross-channel opportunity, there are changes in the ecosystem that allow us to move up-funnel. CTV is a channel that allows you to move up-funnel. Discovery, as a bigger and more fragmented layer, allows you to move up-funnel. So channels and funnels go hand in hand. If you have access to a new channel mix, you now have access to new funnel budgets, and that’s the strategy for that part of the business.
We can take that recommendation engine, targeting capability and measurement framework and move across channels enabling us to go after budgets that sit in different parts of the funnel than what we’ve typically supported. That’s the strategy.
Any plans through the end of the year as it relates to performance?
We’ll continue to add new channels. CTV is a channel that we’re still a little bit underdeveloped. And then, there’s an interesting segmentation of product where GO is completely self-service, and therefore targeted at small- and mid-sized businesses. We’ve even had some large clients taking advantage of that. I think there may be a version of that which will support clients who want more control. But we’ll also have an element of self-service for our XL and large clients.
We’ll be discovering what’s the right balance of self service versus control and what’s the right product interface or MCP setup to enable that.
That’s what we’re thinking about for the second half of the year.
Given your past experience at Dentsu among other career milestones, I’m curious what you think about the agency model today?
I’d start with the bear case theories about agencies going away. These theories are way overdone. Agencies will remain a vital part of this ecosystem.
For forever this ecosystem gets more complex every week and agencies are trusted partners to help tame that complexity. No matter how much great tooling ad tech companies or walled gardens put out, there will always be a need for a business partner that helps with measurement, strategy and planning. So that’s important to say up front.
I think agencies have a very viable model and vital role in the ecosystem. That said, clearly, there’s pressure on the FTE (full-time employee) model. Paying for headcount in an AI-enabled world is antithetical. But clients have to help figure that out, too.
Agencies are more than willing to adapt on how to be compensated on outcomes. Clients and procurement departments need to help by enabling that as well. That’s the crux of the evolution for agencies over the next few years — how to transition that compensation model.
Similar to the agency model, many are discussing the end of enterprise SaaS models with the advancements of AI and LLMs. Any thoughts?
My point of view is similar to how I tackle many other industry theories that pop up. SaaS companies are absolutely viable going forward. But the impact AI has had on SaaS platforms may have removed some of their value.
But I would never go so far as to think that it’s not viable or valuable and AI is going to replace all of this deep expertise and functionality which SaaS platforms provide. It gets down to the data layers in terms of who has access to what. How is it permissioned? How does a workflow operate inside of an enterprise? There’s deep domain expertise that SaaS products have embedded in their functionality. I don’t see AI becoming a sentient capability that’s able to replicate all that.
I realize you’re limited in what you can say regarding the future in that Criteo is a publicly-traded company. But, three years from now, what does Criteo look like?
It looks like our strategies today more fully manifested — so full funnel + cross-channel, self-service, scaling retail media. Imagine what those could be three years from now?
Regarding full funnel, cross-channel, we should be one of the major players in the discovery layer. We should be more built out in the CTV channel. Marketers should be executing campaigns with companies like Criteo on a cross-channel basis, whether through MCP or through self-service interfaces like GO.
It should be much more hands-on and simple to manage that complexity. And retail continues to serve shopping in whatever format or platform it takes with the right balance of organic and paid as the engine of what gets served. The art of retailing will be alive and well. There will just be more surfaces and channels in which it takes place.