Criteo introduces commerce-grade recommendation infrastructure for AI shopping assistants, achieving 60% relevancy gains through transaction data versus product descriptions alone.
Criteo today unveiled technology designed to address a fundamental limitation plaguing AI shopping assistants: their inability to distinguish between products that merely match a query semantically and those consumers actually want to buy.
The global commerce media platform's Agentic Commerce Recommendation Service, announced February 5, demonstrated up to 60% improvement in recommendation relevancy compared to approaches relying solely on product descriptions in internal testing conducted during January 2026. The performance differential stems from what Criteo characterizes as "outcome-based relevancy" - recommendations grounded in actual shopping behavior rather than textual similarity between queries and product catalogs.
According to the company's research published alongside the commercial announcement, most agentic commerce recommendation services currently depend on content embeddings or mainstream large language models including Claude, Llama, and GPT. While these systems excel at semantic understanding, they lack grounding in real shopping patterns and optimization for commercial objectives such as clicks or conversions.
"The real competitive advantage in agentic commerce will come from access to high-quality commerce data at scale," according to Michael Komasinski, Chief Executive Officer of Criteo. The service brings that intelligence into AI-driven shopping experiences through the company's unmatched scale of 720 million daily shoppers, $1 trillion in annual transactions, and 4.5 billion product SKUs.
The distinction between semantic accuracy and outcome-based relevance represents more than technical nuance. Two products may match a user query equally well from a content standpoint - sharing brand, category, and attributes - yet differ significantly in their ability to meet user expectations, drive engagement, or lead to purchase. Traditional approaches optimized solely on retailer catalog data can achieve semantic matching but miss the signal about which products consumers prefer when given similar options.
Amazon deploys generative and agentic AI across shopping platform with 250 million Rufus users throughout 2025, while Microsoft's Copilot integration delivered 53% more purchases within 30 minutes of interaction compared to journeys without the AI assistant. The agentic commerce infrastructure race intensified as platforms recognized that product discovery increasingly occurs through conversational interfaces rather than traditional search.
Criteo's implementation employs a two-stage architecture that first retrieves candidate products through content embeddings optimized on commerce data, then re-ranks those candidates for maximum outcome-based relevancy. The approach differs from single-stage systems that attempt to solve both problems simultaneously.
Product retrieval happens through the company's CLEPR model - Contrastive Language Embedding for Product Retrieval - a 120 million parameter bi-encoder using Multilingual-MiniLM-L12-H384 as backbone. The system has been fine-tuned on organic click data and currently powers keyword-based product recommendations across Criteo's platform.
The model computes separate embeddings for queries and products, enabling efficient K-nearest neighbor search across large product catalogs. Each product receives characterization through brand, category, and full textual title. According to the company's benchmark results, CLEPR outperformed widely-used open-source text encoders from Google, Microsoft, Meta, and Alibaba by 37% in outcome-based relevancy on average while maintaining smaller size and greater efficiency.
Product re-ranking then orders candidate products using both query-recommendation accuracy and commerce-driven scores including sales velocity. The system applies what the company describes as PSales probability - the likelihood of a product being sold once displayed and clicked, calculated as the proportion of sales observed during the previous seven days.
The architecture enables nuanced product recommendations that understand broad shopper intent while supporting both exploratory queries and product-specific searches. The system delivers relevant product recommendations and expands them with complementary items when appropriate, according to the announcement.
Criteo's testing methodology reveals the magnitude of commerce data's impact on recommendation quality. The company evaluated outcome-based relevancy by measuring the ability to identify products actually clicked for given queries from among 400 alternative products. Negative candidates were defined as alternative clicked products from other user searches sharing either the same brand or same category as the target product.
This approach forces models beyond surface-level semantic similarity to capture user intent and product relevance. The normalized reciprocal rank metric measures how likely users are to find what they seek among top results.
When comparing CLEPR against zero-shot text encoders including Google's offerings, the commerce-trained model achieved 37% higher outcome-based relevancy on average. The performance gap widens further when examining the re-ranking stage, where incorporating transaction-derived popularity signals produced 60% uplift in outcome-based relevancy for SKU re-ranking.
The company's research emphasizes that accuracy - whether product descriptions semantically match user queries - differs fundamentally from outcome-based relevance measuring whether recommendations satisfy underlying user needs as evidenced by real behavior. High accuracy can be achieved using retailers' catalog data alone, according to the technical documentation. Outcome-based relevance requires commerce data.
This distinction carries significant implications for the emerging AI shopping assistant ecosystem. As AI shopping adoption accelerates, with 85% of UK consumers planning AI-assisted holiday shopping willing to trust autonomous agents to place orders according to November 2025 research, the quality of underlying recommendation infrastructure becomes increasingly critical.
The service operates through Criteo's Model Context Protocol implementation, directly connecting AI-powered shopping assistants with merchant inventory. The protocol translates consumer shopping queries into curated, transaction-ready product recommendations by applying real-world shopping and purchase signals unavailable through traditional web crawling tactics.
When consumers request products matching their needs, preferences, and budget through an AI shopping assistant, the assistant queries Criteo's service to identify relevant products. The system applies commerce intelligence-powered filtering that considers nuances including product popularity, availability, and user intent. Rather than returning raw catalog data, the service provides curated shortlists of product recommendations.
The AI assistant then reviews Criteo's recommendations, presents results, compares options, and can support add-to-cart or checkout within the agentic experience. This workflow differs from implementations where AI assistants scrape product information directly from retailer websites or rely on publicly available product feeds lacking behavioral signals.
The protocol-based approach addresses what the company frames as an infrastructure gap in AI-assisted shopping. As Target and Walmart bring checkout directly into Google's AI assistant and PayPal teams with Microsoft to enable purchases inside Copilot, the underlying recommendation engines powering these experiences require access to purchase behavior data that platform operators typically cannot access at scale across the broader commerce ecosystem.
Criteo continues testing with a major LLM platform that began in 2025, while expanding validation to additional LLM platforms, retailers, and brands. The company has not disclosed specific platform partners or timeline for general availability beyond the testing phase.
The service arrives as Criteo executes broader strategic positioning within retail media and commerce advertising. The company's retail media business exceeded $250 million in annual revenue for 2024, establishing it as significant player in the fast-growing sector. Fourth quarter 2024 retail media contribution ex-TAC grew 23% at constant currency to reach $90 million.
Recent product developments include auction-based display ads for retail media launched in June 2025, which brought programmatic bidding dynamics to retail environments. The company also announced partnership with DoorDash for multi-year retail media expansion in October 2025 and commerce-driven CTV activation with WPP Media in July 2025.
However, the company faces headwinds including major retail media client relationship changes announced in May 2025, when its largest retail media client discontinued managed services while maintaining technology platform usage under multi-year contract. The shift resulted in expected $25 million negative impact in 2025.
The client changes reflect what CEO Komasinski described as transformation "from largely managed service model to a more scalable self-service platform." This strategic direction aligns with the agentic commerce recommendation service, which provides technology infrastructure rather than labor-intensive service delivery.
The agentic commerce announcement arrives amid fundamental tension between AI shopping assistants and retail media monetization models. Traditional retail media networks generate revenue through sponsored product listings, display advertising on retailer websites, and off-site advertising driving traffic to retailer properties. When transactions occur entirely within AI assistant interfaces, these monetization models require adaptation.
Whether platforms will introduce sponsored placements within AI assistants' product recommendations - and how such advertising would function in conversational interfaces - remains largely unaddressed across the industry. The competitive dynamics intensify as retail media is projected to capture 20% of global advertising revenue by 2030, reaching over $300 billion annually.
The financial incentive creates tension with AI-mediated shopping models that potentially diminish platform control over discovery and purchase journeys. Amazon, controlling more than 25% of U.S. ecommerce, has blocked external AI bots through technical restrictions while developing internal tools like Rufus that gained memory, price tracking and auto-buying capabilities in November 2025.
Walmart Connect bets on AI agents to reshape retail advertising through an advertising assistant announced January 6 that helps brands build, optimize, and troubleshoot campaigns using conversational chat. The retail media network tested advertising formats within its Sparky agent during fall 2025, marking first integration of advertisements into AI shopping experiences.
Criteo's positioning as commerce infrastructure provider rather than direct retailer potentially enables collaboration across competing retail media networks. The company partners with 70% of the top 30 retailers in the U.S. according to May 2025 disclosure, including recent wins such as Dick's Sporting Goods, Endeavour in Australia, and d shopping in Japan.
The technical research accompanying the commercial announcement provides detailed comparison between commerce-optimized models and general-purpose text encoders. Criteo evaluated CLEPR against models from the MTEB leaderboard of Hugging Face, using dataset of user search queries paired with clicked products.
The benchmark assessed accuracy through ROC-AUC metric on human-annotated dataset where binary labels indicate semantic match quality between queries and products. Fine-tuning on commerce data provided moderate accuracy improvement - 3 points compared to better text encoders, 6 points versus the same backbone model without fine-tuning.
However, the accuracy metric measures only semantic relevance rather than outcome-based performance. The outcome-based relevancy evaluation, which measures ability to identify actually-clicked products from alternatives sharing brand or category, showed far more substantial performance differential favoring commerce-trained models.
The research notes that CLEPR solution is trained solely on organic clicks and does not include bias terms associated with sales events. Consequently, the company cannot confidently report global uplift across end-to-end pipeline based on sales conversion metrics at this stage. The independent uplifts for SKU retrieval and re-ranking steps provide confidence in similar global performance for complete agentic commerce recommendation service.
The findings align with broader industry developments in recommendation systems and product discovery optimization. Marketing professionals managing multi-channel campaigns increasingly recognize that semantic similarity alone proves insufficient for driving commercial outcomes at scale.
Future technical development will include deep dives into multimodal embeddings and LLMs for efficient product search, LLMs and fine-tuning techniques for product recommendation explanations, and deep learning-based product re-rankers according to the research roadmap. The company plans continued blog series on enabling technologies for agentic commerce recommendation services.
The announcement occurs against backdrop of skepticism about autonomous shopping systems' commercial prospects. Analysis examining structural challenges to agentic commerce adoption identified retailer incentives against AI intermediation, high ecommerce return rates, and consumer preferences for evaluating options before purchasing as significant barriers.
McKinsey research projects that agentic commerce could orchestrate between $900 billion and $1 trillion in revenue in the U.S. B2C retail market alone by 2030, with global projections reaching $3 trillion to $5 trillion. However, actual adoption remains limited. A 2025 survey found only 12% of consumers had used AI chatbots for product research, with fewer completing purchases through these interfaces.
The low adoption reflects both limited availability of commerce-enabled AI assistants and consumer unfamiliarity with conversational shopping patterns. Past retail technology innovations from QR codes to augmented reality shopping apps have struggled achieving mainstream adoption despite significant investment.
Technical implementation challenges persist around product discovery accuracy, transaction security, returns handling, customer service, and merchant profitability. The revenue sharing arrangements between retailers and AI platform providers remain largely undisclosed across announced partnerships.
The scope of product categories suitable for conversational shopping also remains unclear. Complex purchases requiring extensive research and comparison may not translate well to conversational interfaces, while commodity purchases may not benefit sufficiently from AI assistance to justify implementation costs.
Criteo's emphasis on outcome-based relevancy rather than semantic matching suggests recognition that commerce-grade infrastructure requires more than general-purpose language models can provide. Whether this technical capability translates to consumer adoption and merchant satisfaction represents the market validation test ahead.
The company's Model Context Protocol implementation enables AI assistants to access Criteo's recommendation infrastructure without requiring merchants to modify backend systems. This interoperability approach mirrors developments including OpenAI's Agentic Commerce Protocol co-developed with Stripe and Google's Universal Commerce Protocol announced January 11.
Standards-based integration potentially reduces friction for merchant participation compared to platform-specific implementations. However, the fragmented landscape of competing protocols and platforms creates complexity for retailers evaluating which systems to support.
Who: Criteo, the global commerce media platform trading on NASDAQ under ticker CRTO, introduced new technology for AI shopping assistants. The announcement came from CEO Michael Komasinski and the company's research and development team, targeting LLM platforms, retailers, brands, and AI assistant developers.
What: The Agentic Commerce Recommendation Service provides commerce-grade product recommendations for AI shopping assistants through Model Context Protocol integration. The service demonstrated up to 60% improvement in recommendation relevancy compared to approaches based solely on product descriptions in internal January 2026 testing. The technology employs two-stage architecture with CLEPR model for product retrieval and commerce-driven re-ranking using transaction data from 720 million daily shoppers, $1 trillion annual transactions, and 4.5 billion product SKUs.
When: Criteo announced the service on February 5, 2026, following internal testing during January 2026. The company continues testing with a major LLM platform that began in 2025, while expanding validation to additional platforms, retailers, and brands without disclosed timeline for general availability.
Where: The service operates through Criteo's Model Context Protocol implementation, connecting AI-powered shopping assistants with merchant inventory across the company's network of 17,000 e-commerce sites, 200 global retail partners, and thousands of open web publishers. The company partners with 70% of the top 30 retailers in the U.S. and operates globally across Americas, EMEA, and APAC regions.
Why: AI shopping assistants currently rely on content embeddings or mainstream large language models that lack grounding in real shopping behavior and are not optimized for commercial objectives. The service addresses fundamental limitation between semantic accuracy and outcome-based relevance by incorporating actual purchase signals, product popularity, and shopping behavior patterns. The announcement positions Criteo within intensifying competition as agentic commerce is projected to orchestrate $900 billion to $1 trillion in U.S. B2C retail revenue by 2030, while retail media networks face structural pressure from AI shopping agents potentially disintermediating traditional advertising relationships.