Hopefully everyone feels good coming out of Thanksgiving and America’s November shopaganza, BFCM, the Cyber Five, or whatever you call it. May your Q5’s be merry and bright and full of attributable sales.
Now back to business. Which is to say, Criteo’s entrance to the AI arena.
On Tuesday, Criteo debuted new AI tech and pilot programs to a group of reporters, which included one back-end shopper data partnership with an unnamed LLM.
The idea of a product
The ad tech company’s new LLM integration is built off of Criteo’s previously released “Model Context Protocol” or MCP, which is the integration point for agentic AI solutions or LLMs to plug into Criteo’s data.
With this new integration engineers are working together to see whether plugging into Criteo’s data via its MCP generates superior results for shopping-related prompts, Criteo CEO Michael Komasinski said. The individual using the LLM doesn’t need to know Criteo or ever see the name, but Criteo’s product recommendation system would contribute to the LLM’s suggested products.
Whether this partnership will ever go live is unknown. As is the potential business model. How Criteo might earn revenue in this arrangement hasn’t been part of the discussion yet, Komasinski acknowledged.
Criteo might license data, akin to how large media companies have signed licensing deals with LLMs. Perhaps it will be more of a traditional advertising relationship with a cost per click, or an affiliate or performance marketing deal.
“I think it gets a lot more interesting in the context of a native ad solution,” he said of a potential LLM partnership. But Criteo might also charge a fee per crawl or query by the LLM, he added.
“We’re purely at the stage of: ‘Does it work?’ and ‘Does it make the results better?’” he said.
Criteo’s ante
In addition to figuring out the monetization part, Criteo must also establish its credentials as a data player in the AI world.
If the company can prove that its data set would be a powerful training system for an LLM or other AI service, Criteo could have a major stake in AI.
“It is data sets like (Criteo’s) that come from sort of advertising recommendation systems that are a big part of what LLMs will need if they want to get into the commerce game,” Komasinski said.
For one thing, he noted, Criteo has a daily active user count of about 720 million; this compares to ChatGPT at about 800 million users. Criteo “observed” $1 trillion in gross merchandise sales last year, he said; Amazon registered a mere $800 billion. And Shopify doesn’t publish its total SKU count, he added, but Criteo believes it is “several billions,” and in the ballpark of Criteo’s 4.5 billion SKUs.
Take those metrics with a cup of salt, though.
ChatGPT’s users are ChatGPT users. When Komasinski cited Criteo’s DAUs, he meant the number of individuals the platform targets through online advertising. Throughout the rest of the presentation, Criteo execs refer to its “users” as agency, retailer or advertiser clients who actually log into Criteo systems, of which there are not hundreds of millions.
Likewise, Amazon’s transaction data is Amazon transaction data. While Criteo might “observe” transactions it sends to retailer or merchant sites, it owns none of that data.
Still, the point is valid that there are very few vast archives of ecommerce data out there. And the likes of Amazon, Walmart and Shopify are not open to working with LLM operators to help improve their services. In fact, Amazon and Shopify block most LLM scraper bots from their sites by default, while Walmart heavily restricts any AI bot that accesses its site.
There must be an Agent
Komasinski did announce one product that actually is launching. Audience Agent is a way to set up campaigns and analytics using natural language prompts without the need for data science or analytics expertise.
The big platforms now routinely boast of “one-click” campaign setups for their AI products. Criteo’s Audience Agent is a six-click service, Parsons said, from generating an audience to generating creative and activating a campaign.
The outputs for Audience Agent segment aren’t exactly revolutionary. For instance, prompting for audiences that are potential buyers for a hiking backpack generated segments like “in-market for backpacks” and “in-market for camping gear.”
Wouldn’t it be more impressive if the results were less intuitive, like targeting woman who buy organic snack bars or that an advertiser should target on rainy days?
“We call those unexpected correlations,” Parsons told AdExchanger after the presentation. “We do see those more often.”
However, he added that there is a more complicated trade-off when the question becomes about how to most effectively buy ads to generate sales. By layering in the extra targeting and refining the audience to focus on woman who buy organic snacks, and on rainy days, the media is more expensive and the overall audience reach is much diminished.
For Criteo as a retargeter, the metric was always ROAS for a click-based ad campaign, Parsons said. Nowadays, he said more advertisers think in terms of incremental ROAS, and whether a campaign generates enough new customers and overall sales to justify the media budget.
Which is also why Criteo has been pushing further upper funnel, he said. For Criteo advertisers that use its dynamic allocation product Commerce GO! (the shouting is part of the brand) to control campaign optimization as they might with Google’s Performance Max or Meta’s Advantage+ Shopping Campaigns – Meta is now 35% of that spend.
The addition of platforms like Meta and TikTok, where Criteo will now extend its open programmatic campaigns via its own ad-buying seat on those platforms, seems like a real philosophical change. But it’s not, Parsons said.
For Criteo, he said, the focus is always on generating a purchase. And those are platforms where people spend their time and think about shopping. There are measurement challenges and Criteo still must justify that its spend on those channels is generating incremental ROAS, even when brands have their own separate social ad team running Meta campaigns.
For most advertisers, though, it’s seen as a simplification for a platform to span open web, social media, CTV and other channels, as long as they see the purchases coming through.
“A point which I think about a lot,” Parsons said, “is how far can we go into that upper funnel with those unexpected correlations and still prove that we’re a reasonably good alternative to Google and Meta? Time will tell.”
