Moloco’s CARA
Our Compound Ad Recommendation Architecture for Moloco Commerce Media
Overview
Retail advertising is deceptively complex. A single platform can have millions of shoppers, millions of catalog items, tens of thousands of sellers, and dozens of ad placements across search, product, and category pages. Every request is expected to be resolved in tens of milliseconds, and the result has to serve the shopper, the advertiser, and the retailer at once — three objectives that often pull in different directions.
What makes commerce uniquely powerful for advertising is the closed loop. Search, impression, click, add-to-cart, and purchase all happen inside one environment, so the signal is rich and conversions are deterministically attributed rather than modeled across platforms. At the same time, every retailer is its own world, with its own catalog schema, item taxonomy, page layouts, shopper base, and business goals.
Moloco Commerce Media’s CARA, Compound Ad Recommendation Architecture, understands the complexity of commerce verticals and orchestrates multiple models across our retailers, covering the six technical domains: Signals, Ad Personalization, Bidding, Ad Serving, Demand Activation, and Campaign Automation to drive outcomes for retailers, advertisers, and shoppers. Moloco Commerce Media is built to grow our customers’ ad revenue as a percent of GMV, or A2G. CARA is the intelligence layer behind A2G growth, where advertiser success, shopper relevance, and retailer monetization improve together.
How it works
.png)
CARA’s Six Technical Domains for Moloco Commerce Media
CARA ingests conversion events, catalog data, search queries, clicks, and impressions for each retailer platform, all first-party and closed-loop, and normalizes them into a unified intelligence layer. For example, product catalog attributes are encoded to embeddings that capture item-to-item relations and semantic matching without relying on manual taxonomy, keyword mapping, or category tagging. Every retailer has its own catalog schema, products go out of stock or change prices within the same day, and the same product can be sold by more than one seller. A precise item representation allows personalization, relevance, and retrieval to work across the systems: so when the signal layer understands the catalog, every downstream domain
gets sharper.
For every ad slot on search, product, and category pages, CARA personalizes which items the shopper sees. Our models based on deep-learning architecture ingest signals from multiple sources simultaneously to form dense representations of shopper behavior and product fit, then predict click and conversion probabilities for each available ad candidate. Because items are represented as embeddings rather than by click history alone, a new item can inherit relevance signals from semantically similar catalog items — so sellers can promote items with little or no history, and new shoppers see ads aligned to their search and discovery patterns from the start.
In commerce advertising, a relevant impression is inventory worth monetizing rather than declining by default. CARA's bidding is designed around that principle: it applies closed-loop conversion data and domain-specific orchestration to optimize against three objectives at once — relevance for the shopper, ROAS and budget pacing for the advertiser, and inventory fill for the retailer's A2G. CARA also detects shopper demand shifts — a promotional event can reshape shopper behavior within hours — and adapts bids and pacing in real time to capture the surge.
Ad inventory is the retailer's owned surface — search results, product detail pages, category pages — and CARA governs ad delivery across those inventories and ad formats. A product-detail slot has a different value than a search slot, so CARA quantifies each placement's impact on auction density, ad exposure rate, and expected revenue uplift to monetize available inventory without degrading the shopper experience. Each new placement strengthens the whole system, because it yields richer signals about how shoppers behave.
One of the hardest problems in commerce media is demand activation: turning sellers into advertisers. Sellers hesitate for concrete reasons — no proof of performance, fear of margin erosion, or a perception of advertising as a complex, black-box investment. Retailers, in turn, face sales motions that don't scale. CARA addresses this with Demand Agent, an agentic automation layer that identifies high-potential sellers, generates ready-to-launch campaigns with recommended settings based on catalog and platform data, and triggers outreach at key lifecycle moments. This scales advertiser activation without heavy operational overhead. Advertisers can drive more sales through improved campaign performance, while the retailer preserves the shopping experience through ad relevancy.
Sellers and brands define success metrics (ROAS target, sales goal, budget) and campaign setup is simplified by their chosen outcome, along with intelligent guidance on item selection and bidding. Agents continuously monitor live campaigns in real time, spotting underspend or targeting gaps, and surfacing optimization recommendations. The system optimizes for multiple goal types — Optimize ROAS, MaxSales, Fixed CPC, pay-per-order—each with its own pacing logic, so sellers with limited advertising expertise can still run effective campaigns. Real-time dashboards provide advertisers insights and transparency on campaign performance.
How the system compounds
CARA is not a set of independent features, but a compound system where each domain amplifies the others. Better signal ingestion produces more precise representations of shoppers and items, which sharpen personalization, retrieval, and bidding. Better bidding fills more high-quality opportunities while holding advertiser ROAS and shopper relevance. Better outcomes generate performance proof that makes more sellers willing to advertise and gives retailers confidence to expand supply.
As more sellers activate, auction density increases. With more eligible advertisers competing for each slot, the system selects from a deeper candidate set, improves fill rate, and serves more relevant ads. Those impressions generate clicks, conversions, budget signals, and catalog-level learning that feed back into CARA's models and agents, making the next decision sharper.
The result is compounding A2G growth: every new campaign, placement, item, shopper interaction, and advertiser outcome strengthens the intelligence layer behind the next decision. The platform moves from manually operating an ad business to running a self-improving commerce media system where advertiser success, shopper relevance, and retailer monetization improve together.
