Ed Dinichert (Criteo): "We want our Commerce Graph to become an intelligence layer accessible everywhere – from CTV to ChatGPT – and through any media buyer's tools"

Ed Dinichert (Criteo): The goal was to provide an update on the ongoing developments around OpenAI and advertising, with a particular focus on Criteo’s initiatives. We are already working with more than 2,000 advertisers on their ChatGPT advertising campaigns, particularly in the United States, the United Kingdom and Australia.

It was also an opportunity to remind people of everything Criteo can bring to an environment like this. We often hear the same comment: “OpenAI is launching its self-service tool, so why continue working with a partner like Criteo? Eventually, OpenAI will do everything itself.”

I think that is an oversimplification. The examples of Google and Meta show precisely the opposite: there is always additional value to be created on top of the platforms.

Exactly. The common thread running through our strategy is quite simple: going back to basics. Everything starts with insights. Good insights enable better activation decisions. Those activations then produce measurable outcomes, which in turn generate new insights. It is a virtuous cycle.

Today, our industry has become extremely fragmented. Everyone talks about measurement, but before you can measure anything, you first need to generate relevant insights. In my view, the companies that succeed will be those capable of connecting all this data, turning it into actionable insights, informing their activation decisions and accurately measuring the results.

At Criteo, we have already integrated data from OpenAI into our internal platform. Our focus now is on making these insights accessible everywhere, particularly through MCP, so they can be used directly from the conversational assistants and platforms used by media buyers.

In practical terms, it means mapping user behavior and comparing what happens with and without exposure to ads served in ChatGPT. We can do this thanks to our Commerce Graph, which currently includes nearly 5 billion SKUs, captures signals from approximately 740 million daily users and observes more than $3 billion in transactions every day.

This allows us to identify changes that we could not previously observe. For example, we are beginning to measure the impact of a ChatGPT campaign on brand consideration and, eventually, probably on loyalty. These are the kinds of metrics we are currently developing.

Not directly yet. However, we have developed conversion proxies that already allow us to estimate that impact. The goal is to understand the role ChatGPT actually plays in the customer journey.

Most advertisers are still in the experimentation phase. The question they ask us is ultimately quite simple: where should this new channel sit within their media mix? Is it next-generation SEO, GEO—Generative Engine Optimization—a performance channel or a brand-building medium?

ChatGPT is primarily a consideration channel positioned in the middle of the funnel.

Yes, that is clearly the direction we want to take. Today, we cannot yet use signals from our Commerce Graph to directly target campaigns on OpenAI.

We are not yet operating within a fully deterministic model where an advertiser could send us a brief, have us build a proprietary audience from our data and immediately activate it, as we already do on Meta, TikTok and other platforms. That will come at a later stage.

Previously, brands had to optimize their search visibility around known queries. Today, they need to anticipate the hundreds or even thousands of ways users might phrase a request.

The better you are at anticipating these contexts, the more opportunities you create for the brand to appear. And the better you are at producing creative variations tailored to those different contexts, the stronger your performance will be.

Not systematically, and it is far too early to generalize. However, we have already seen situations in which GEO performance increased by 50% to 80% following certain ChatGPT activations.

That may simply mean that the language model did not yet know enough about a particular brand or entity, and that the campaign helped make it more visible within the model’s knowledge universe.

I think the browsing experience is going to change profoundly. The widespread adoption of large language models will transform the interfaces of most websites. Over time, we will navigate much less through traditional menus and much more through assistants capable of understanding user intent.

In that context, the challenge will be to connect these new interactions with product catalogs, SKUs, marketing data, campaigns and insights. It is therefore an extremely relevant bet, although the real question is how much volume this inventory will ultimately represent. At this stage, no one knows for certain.

In any case, the Albertsons deal is a natural extension of our strategy: bringing together these new data sources and making them actionable across channels and throughout the funnel, including within retail media.

Yes. We currently work with Meta, TikTok and OpenAI. In reality, far more players are open than people might assume.

Yes. We have already announced our integration with PubMatic and hope to finalize integrations with other SSPs soon.

The idea is simple: to make our Commerce Graph available within the leading programmatic environments. Today, we provide targeting. Tomorrow, we will gradually add insights and then measurement.

Because the market is sending us a very clear signal. In connected TV in particular, premium publishers are primarily looking for simplicity. No one wants to add another SSP to their tech stack. On the contrary, buyers want to reduce the number of intermediaries.

I have no interest in asking Disney+, Netflix or major broadcasters to overhaul their entire infrastructure just to work with us.

Curation within third-party SSPs is our entry point into the CTV market.

However, all of them are interested in our Commerce Graph, targeting capabilities and measurement solutions. We therefore decided to bring that intelligence directly into the environments where buyers already operate.

It is much simpler for the ecosystem as a whole, and it is also an excellent way for Criteo to accelerate its development in the CTV market.

His arrival has been tremendously helpful. Julien has an in-depth understanding of the expectations of publishers, agencies and advertisers. Thanks to him, we are asking the right questions much earlier in the process: Which SSPs should we prioritize? Which video partners should we integrate? How should we structure our roadmap?

In just a few months, he has helped us prioritize our development work much more effectively. He brings a very detailed understanding of the CTV ecosystem.

We conducted a number of demonstrations for agencies, which were able to see that they could query our platform directly using natural language.

For example: “Show me the best customers for this category.” “Which audience segments perform best?” “Which products are performing best?” All this information becomes accessible without users having to navigate a complex interface.

This represents a profound shift. Our platform is no longer simply a tool reserved for media traders. It becomes a genuine intelligence layer that can be accessed through an assistant.

Eventually, the development of MCP servers by different advertising platforms could provide a solution to media fragmentation. It could make it possible to orchestrate much more comprehensive campaigns.

We have not yet made all our interfaces available to agencies on a self-service basis. But our ambition is clear: to make our Commerce Graph and measurement capabilities accessible everywhere. We want to become an intelligence layer that anyone can integrate into their own tools.

I believe they do. In fact, I feel quite strongly about it. Every major technological transformation leads people to predict the disappearance of operational roles.

We heard exactly the same arguments with ad networks. Then with mobile. Then with RTB. Then with programmatic. Today, it is agentic AI.

Every time, people predict that everything will become automated. The reality is far more nuanced. There will always be companies that lack the time, resources or expertise and prefer to delegate. Managed services will therefore continue to exist.

That is probably the real issue. Yes, some organizations will likely have fewer employees. But that does not mean they will immediately have the tools required to replace all human work.

Quite the opposite. During the transition period, many companies will need external partners to help them manage the transformation. I even suspect that demand for support could increase in the short term.

Today, many teams spend more time building tools than actually using them. Some agencies are already reorganizing their workforces.

They are expanding their product and engineering teams while gradually reducing certain operational roles. But this shift will take time. Our brains like simple scenarios: “AI is coming, so everything will be automated.” In reality, transformations are always far more gradual.

I do not think the real question is whether users will continue to visit chatgpt.com. The real question is: Where will these language models be deployed?

Today, we still think of ChatGPT primarily as a website or an app. I do not believe that will be the case tomorrow. Eventually, these models will be embedded everywhere.

They will be present across e-commerce websites, apps, customer service platforms and professional tools. Companies will use intelligence from OpenAI, Anthropic or other providers without necessarily sending users to those companies’ respective interfaces.

That is exactly what is beginning to happen. To my knowledge, apart from Amazon, very few retailers are currently developing their own language models. Most prefer to rely on existing infrastructure.

Exactly. I often compare it to Shopify. Shopify is not the merchant’s website. It is the infrastructure that allows merchants to build their businesses. Tomorrow, language models will probably play a similar role. They will become a technology layer embedded across a wide variety of services.

Exactly. That is when the potential moves to an entirely different scale.

If these assistants become distributed infrastructure, as Shopify did for e-commerce, their advertising market will be far larger than we currently imagine. That is probably the right way to look at it.