It’s Not Enough to Have LLMs. Commerce Needs LCMs.

‍Search changed how people find products. Mobile changed where they shop. AI agents could change how much of the journey they delegate. Commerce needs models that understand what happens next.

Think about the last time you bought something online. You searched, scanned a list of products, compared prices and availability, opened a few tabs, and made the decision yourself. For years, commerce technology has helped people navigate those steps. The next shift may let AI take on more of the work.

Every new interface changes commerce

Past technology shifts show how much an interface can change commerce. In the U.S., the number of people using search engines on an average day rose from about 38 million in June 2004 to 59 million in September 2005, a 55% increase in roughly fifteen months, according to Pew Research Center.

Smartphones followed a longer adoption curve: U.S. smartphone ownership rose from 35% in 2011 to 91% by 2024. By the 2024 holiday season, smartphones accounted for 54.5% of online transactions tracked by Adobe.

AI-assisted commerce is showing early signs of adoption. Shopify reported that AI-referred sessions to its stores grew more than eightfold year over year in Q1 2026, while orders from AI referrals grew nearly thirteenfold. Amazon reported that its shopping assistant was used by more than 300 million customers in 2025 and helped generate nearly $12 billion in incremental annualized sales.

These measures come from different companies and reflect different kinds of activity, but they point to AI becoming part of how people discover and buy products.

The next step is agentic commerce: systems that can help compare options and, increasingly, take actions for shoppers. We should be careful about predicting exactly how quickly people will delegate purchases. Discovering products with AI and authorizing an agent to buy are different steps. Still, the interface can reach people through the devices and platforms they already use.

McKinsey estimates that by 2030, agentic commerce could orchestrate $900 billion to $1 trillion in U.S. B2C retail revenue, and $3 trillion to $5 trillion globally. Those figures are forecasts of potential revenue, not measures of current adoption.

Commerce needs intelligence beyond language

As these interfaces change, commerce systems need to do more than understand a request. They need to find relevant products, account for price and availability, understand shopper intent, weigh business goals and learn from what happens next.

That is why commerce needs its own kind of intelligence: Large Commerce Models, or LCMs.

Large language models are trained to predict the next token. A commerce model has a different task: to understand the signals that lead to a transaction, predict likely outcomes and help a business make a  better decision in response.

Those signals are not just words. They include product attributes, searches, clicks, purchases, prices, inventory, shopper behavior, advertiser demand and the results of previous decisions.

A product can be relevant but unavailable. A shopper can move from browsing to buying. A promotion can increase sales while reducing margin. Commerce intelligence has to understand these relationships in the context of the business.

This is more than adding product descriptions to a general-purpose model. A commerce model needs to learn from commerce behavior and transaction outcomes, then connect that understanding to decisions such as retrieval, ranking, recommendations and monetization.

Language models can provide a natural interface. Commerce models can provide the intelligence behind what the system does.

Better product understanding can improve discovery

Early examples of conversational shopping suggest that better product understanding can matter.

Mastercard reported that initial Shopping Muse tests at Michael Kors produced conversion rates around 15–20% higher than traditional search queries. In a separate Bergzeit case study, shoppers who engaged with Shopping Muse converted at 3.4 times the average rate.

That second figure describes engaged shoppers rather than a randomized comparison, but both results point to the value of helping people express what they want and find relevant products.

Building a model for commerce decisions

At Topsort, our work in retail media has given us a close view of the decisions behind commerce: what to show, how to rank it, how to account for demand and availability, and how to learn from outcomes.

We recently introduced T-Brain, our enterprise AI foundation for commerce, powered by Theo, our Large Commerce Model. Theo is built for real-time commerce decisions, with sub-five-millisecond inference.

In our model evaluation, fine-tuning improved click prediction across all seven marketplaces tested, with an average gain of 5.4 AUC points. Purchase-after-click prediction improved by an average of 10.6 AUC points.

Separate retrieval and ranking evaluations covered catalogs from 130 to 894,000 products. On one 130-product catalog, the share of first suggestions judged right rose from 46% to 62%.

These are prediction and retrieval evaluation results, not claims of conversion or revenue lift. They show why commerce models need to adapt to the environment where they operate: marketplaces differ in their catalogs, shopper behavior and patterns of demand.

Speed matters, too. A model that informs search results, rankings or an auction has to respond within the time available for that decision. Specialized commerce intelligence can help make those decisions in real time, using the signals that matter to the retailer and shopper.

From understanding intent to deciding what happens next

LLMs and LCMs are complementary. Language models make it easier to express intent. Commerce models help systems understand the products, behaviors and business outcomes involved. As AI agents become another way to shop, commerce will need both.

Search helped people find information. Mobile put the store in their pocket. Commerce models will help businesses understand what shoppers are trying to do—and decide what should happen next.

The next model commerce needs will predict more than the next word. It will help predict—and shape—the next transaction.

Frequently Asked Questions

What is a Large Commerce Model (LCM)?

A Large Commerce Model is designed to learn from products, shopper behavior, context and transaction outcomes. It uses that understanding to support decisions such as which products to retrieve, how to rank them and how likely a shopper is to click or purchase.

How is an LCM different from a large language model (LLM)?

LLMs are built around understanding and generating language. LCMs focus on commerce signals, including impressions, clicks, purchases, product relationships and pricing. The distinction lies in what the model learns from and what it is optimized to predict: language capability helps interpret a request, while commerce-specific learning helps determine which products and actions are relevant.

How do T-Brain and Theo fit together?

T-Brain is Topsort’s enterprise AI foundation for commerce, powered by Theo, its Large Commerce Model. Theo brings commerce-specific learning to tasks such as retrieval, prediction and ranking, with adaptation to each business’s catalog and shopper behavior.