Criteo Opens Commerce Data Assets via Model Context Protocols

Full analysis

Criteo is taking the shopping-behavior data from its 17,000 commerce customers and exposing it through Model Context Protocol, the emerging standard that lets AI systems pull in outside data without a custom build. Todd Parsons is pitching it two ways at once: a new revenue line from selling data access, and the plumbing that turns AI-driven product discovery into a performance channel you can actually measure and repeat.

Reversibility: Type 2. Opening an MCP endpoint is cheap to launch and cheap to walk back. The strategic bet underneath it, repositioning Criteo as commerce infrastructure rather than a retargeting shop, is Type 1 and much harder to reverse once analysts and customers file you under one label.

What's actually being decided: Not "should Criteo publish an API." It's whether Criteo can matter in the AI-discovery stack before that stack hardens around data somebody else owns.

Forcing function: MCP adoption is happening now, driven by developers, not procurement. First-mover positioning is the whole game or there's no point.

The Market Analyst. Criteo's stock has been beaten down because the market thinks retail media incumbents eat its lunch. This is a TAM reframe: stop selling CPMs, start selling data access, and suddenly you're compared to LiveRamp and the clean-room crowd instead of to Amazon Ads. In plain terms, Criteo wants to be seen as a data utility, not an ad vendor, because utilities get valued higher. The tell of whether it's real is distribution. If Snowflake or Databricks shows up as a channel for this, there's enterprise pull. If it stays a Criteo-hosted endpoint nobody routes through, it's a sentiment pop with no earnings behind it.

The Skeptic. Seventeen thousand customers is a scale anchor, not a usage number. The question is how many AI builders query this endpoint six months out, and the honest base rate is low. Enterprise AI teams prioritize the first-party data they already own before they reach for a retargeter's behavioral graph. Amazon has the same purchase signal natively; Google has the reach. "Repeatable performance-marketing channel" is Parsons selling a future, not describing a present. For a non-specialist: Criteo is offering to sell you a view of how shoppers behave, in a market where the two biggest buyers of that pitch already generate it themselves.

The Operator. The real unlock is that a mid-market team gets behavioral signal without staffing a data-partnership function. That's genuinely useful. But MCP is young, and the first 90 days will surface governance friction before a single query hits production. Data governance teams stall on schema and PII handling for weeks. The quiet killer is freshness. Criteo's signal is worth something because it's near-real-time, and the moment an enterprise drops a caching layer in front of it, the advantage degrades and nobody notices until the campaign underperforms. Explained simply: the data is only valuable while it's fresh, and most buyers will accidentally make it stale.

The Strategist. The two-year logic is sound. A commerce data layer becomes a moat only if it's embedded in third-party AI workflows before those workflows lock onto a competitor's signals. Using MCP as the delivery mechanism is smart positioning, because it frames Criteo as infrastructure that other tools plug into rather than another ad platform fighting for budget. The catch is the same protocol that lets Criteo distribute widely lets anyone else do it too. A clean-room player like InfoSum or an exchange like Index can stand up a comparable data surface over the same standard. Being early buys a head start, not a wall.

Where they part ways:

1. Is the interface a moat or a commodity? The Strategist says standardizing on MCP is how you get embedded everywhere. The Market Analyst and Skeptic say a standard interface means you compete on volume, and volume against Amazon is a race Criteo loses. Both can't be right. If the interface is trivial to copy, then the only durable asset is the data's exclusivity, and Criteo's shopping signal is not exclusive.

2. Does demand exist, or is this developer hype? The Operator sees real mid-market appetite. The Skeptic sees MCP enthusiasm from developers that hasn't crossed into enterprise deployment, and buyers who prefer their own first-party data. This is the crux and it's testable: query volume in six months settles it.

What it hinges on: Two facts. First, whether AI builders actually value commerce behavioral signal over data they already own. Second, whether Criteo's signal is differentiated enough that copying the MCP interface doesn't collapse it into a commodity. If both go against Criteo, this is a sentiment lift with unchanged two-year fundamentals. The council leans skeptical on near-term revenue and cautiously positive on the positioning logic. Before reading this as a turnaround, watch for a named enterprise distribution partner and any disclosed utilization. Neither exists yet.

Prediction: Criteo will not name a major enterprise data-platform distribution partner (Snowflake, Databricks, or a comparable cloud/clean-room vendor) for its commerce MCP offering by its Q4 2026 earnings report in early 2027, and will describe adoption in customer-count and access-availability terms rather than query volume or attributable revenue.

Confidence: Medium — MCP enterprise deployment lags developer hype, and the pitch outruns the demand.

Why: The signal in this story is that Parsons is selling repeatability and openness, not usage, and the summary anchors on 17,000 customers rather than anyone actually querying the endpoint. That's the pattern of a positioning launch: you lead with the interface and the install base because the utilization isn't there yet. The mechanism that keeps a real partner off the stage is that enterprise AI teams reach for their own first-party data first, and Amazon and Google already generate this purchase signal natively, so there's weak pull to route commerce queries through a third party. The opposite outcome, a marquee Snowflake or Databricks deal plus disclosed query volume, would require demand to have crossed from developer enthusiasm into procurement in under two quarters, which almost never happens with a protocol this young. Availability terms let Criteo report progress without proving demand, which is exactly what a company does when the demand is still forming.

Revisit by 2027-03-15: We're right if Criteo's Q4 2026 report and surrounding commentary tout the MCP launch using access, customer-count, or partner-count framing with no named enterprise cloud/clean-room distribution partner and no query-volume or attributable-revenue figure. We're wrong if Criteo names such a partner or discloses concrete usage or revenue tied to the commerce MCP endpoint.

The positioning move is defensible. It's the demand that has to show up, and nothing in Parsons' pitch proves it has.