What Are Large Language Models? How LLMs Work and Where They’re Used

A large language model (LLM) is an AI system that learns statistical patterns from huge amounts of text and other data, and generates output by predicting the next token in a sequence.

Statista calls AI the fastest-adopted technology in history, and awareness is widespread. A 2026 Pew Research Center survey found that 96% of US adults know at least a little about AI, while half that amount say they’ve heard a lot about it.

For ecommerce businesses, LLMs now sit inside search, service, merchandising, and software workflows.

What are large language models?

A large language model is an AI model trained on massive amounts of data to learn patterns in language. The model can generate, summarize, translate, and analyze text, as well as perform other natural language processing (NLP) tasks. The “large” partly refers to its huge number of parameters, that is, the numerical values the model adjusts during training as it learns which patterns help it make better predictions.

The global LLM market size was $7.77 billion in 2025, according to Precedence Research, with chatbots and virtual assistants accounting for the largest application share.

Examples of large language models

A handful of model families account for much of today’s LLM usage, although the names can get confusing. ChatGPT, for example, is a product rather than a model family—OpenAI’s GPT models provide the supporting intelligence, and the particular model serving a ChatGPT conversation can change.

The market looks different depending on what you measure. In Counterpoint Research’s global estimates, Anthropic accounted for 31.4% of LLM revenue in the first quarter of 2026, followed by OpenAI at 29%, Google at 12.1%, and Meta at 1.4%, among others. 

What does open source mean for an LLM?

You’ll often see LLMs divided into open-source and closed-source models. But “open-weight” is often the more accurate term. An open-weight model makes its trained parameters available, which can let developers download it, run it on their own infrastructure, and potentially fine-tune it. 

That doesn’t necessarily mean its training data, training code, or complete development process is public. Llama is a good example. Meta distributes its model weights, but users must accept Meta’s license and comply with its conditions. 

Mistral, meanwhile, takes several approaches. Some models carry the permissive Apache 2.0 license, while others use different licenses or remain commercial.

How large language models work

At a high level, modern generative LLMs use transformer neural networks to learn statistical patterns from massive amounts of training data, then apply those patterns to predict and generate sequences of tokens. 

There are two phases in the process: training, when the model’s parameters are adjusted as it learns, and inference, when it applies those parameters to a new prompt.

Data

LLM development starts with data, usually an enormous collection of text, code, and, increasingly, other media. The exact mix varies by model and provider. 

OpenAI, for example, says its foundation models are developed primarily using publicly available information, third-party data partnerships, and information supplied or generated by users, human trainers, and researchers. As of September 2026, development datasets contain trillions of tokens across text, images, audio, and audiovisual material. 

LLM pretraining commonly uses self-supervised learning, where the training data supplies its own learning signal. A causal language model, for example, can learn by repeatedly trying to predict the next token in a sequence.

More data can help, but volume alone doesn’t determine model quality. DeepMind’s Chinchilla research showed that a 70-billion-parameter model trained on substantially more data could outperform the much larger 280-billion-parameter Gopher using the same compute budget.

Architecture

Many modern LLMs are built on the transformer, a neural network architecture introduced by Google researchers in the 2017 paper Attention Is All You Need.

The key idea is self-attention. Say a copy editor draws invisible lines between words in a sentence, with thicker lines connecting the words that are the most relevant to one another. 

So, in “she deposited cash at the bank,” cash and deposited strengthen the financial meaning of bank. In “she sat on the bank of the river,” river pulls it in another direction. 

The transformer performs a mathematical version of that process, assigning different weights to relationships between tokens so each one can be interpreted in context. This made transformers a lot easier to train at scale than earlier recurrent architectures since many of those relationships could be calculated in parallel.

The original transformer was built for machine translation. Now, variants of the same architecture underpin model families including GPT, Claude, Gemini, and Llama.

Training

In training, text is broken down into tokens, which are the units an LLM processes. A token might be a whole word, part of a word, punctuation, or even a character, and different models can tokenize the same sentence differently. OpenAI’s tokenizer guide describes them as the basic units models receive and generate. 

At the start of pretraining, the model’s parameters, the numerical weights that shape its calculations, have not yet learned useful language patterns. 

Think of training like calibrating a massive machine with billions of adjustable settings. A sequence of tokens goes in, the model predicts what should come next, and that prediction is checked against the actual next token. The difference produces an error score called loss. 

That error is then used to make tiny adjustments across the model’s parameters before the process runs again. After enormous numbers of these corrections, the model becomes much better at predicting language patterns.

Fine-tuning and maintenance

An LLM with pretraining has a broader language capability. Once it’s fine-tuned, developers can test it on a smaller, more targeted dataset. OpenAI, for example, currently supports supervised, preference-based, and reinforcement fine-tuning methods for supported models.

One post-training technique is reinforcement learning from human feedback (RLHF). In OpenAI’s 2022 InstructGPT work, people first demonstrated desirable answers and then ranked different model responses. Those preferences were used to train a reward model, which supplied feedback for further model training. The result was that human evaluators preferred outputs from a 1.3-billion-parameter InstructGPT model over those from the much larger 175-billion-parameter GPT-3 on the study’s prompt distribution.

RLHF is still used, but it’s only one part of modern post-training. Anthropic, for example, says Claude Sonnet 4.5 underwent substantial fine-tuning using techniques including reinforcement learning from both human and AI feedback.

Prompting is different from fine-tuning. While prompting changes the context for an interaction, fine-tuning changes the model itself by adjusting its parameters. 

The benefits of large language models

An LLM can handle a wide range of tasks that involve a lot of language, which explains their rapid adoption. In a 2025 Shopify survey of store owners, 75% said they were already using AI tools.* Common uses for AI in ecommerce include content generation, data analysis, and store operations.

A single model can cover tasks like summarizing a report, writing code, or answering customer questions, often through the same plain-language interface. You can save time and scale your business once you connect governed data and approved workflows:

• A single model can handle many kinds of work. LLMs can answer questions, produce and rewrite copy, translate text, classify information, extract data, summarize documents, assist with code, and support research. Microsoft’s current LLM guidance lists reasoning, content generation, summarization, translation, code generation, classification, extraction, and multimodal understanding among the core capabilities.
• You canwork with software in plain language. LLMs let you give software instructions in everyday language. OpenAI’s 2025 State of Enterprise AI report found that 75% of surveyed users said AI enabled them to complete tasks they previously couldn’t, including programming support and technical troubleshooting.
• LLMs can cut time spent on language-heavy tasks. In a randomized study of more than 6,000 knowledge workers published in 2025, employees who used Copilot integrated with Outlook were able to spend 12% less time on emails each week. 
• The models can be adapted to specific tasks and data. A general-purpose model can be steered with prompts, fine-tuned on task-specific examples, or connected to external information through retrieval-augmented generation.
• The same capability can scale across large workloads. Once an LLM is integrated into a workflow, it can be applied across large numbers of documents, products, customer conversations, or other inputs. 

The limitations of large language models

LLMs can produce fluent work, but that very fluency can make their failures hard to spot:

• LLMs can produce plausible answers that are wrong. These errors are commonly called hallucinations, or “confabulations” in National Institute of Standards and Technology terminology. While they’ve improved considerably, Vectara’s May 2026 hallucination benchmark, as of May 2026, measured rates of 3.1% for GPT-5.4 Nano, 3.3% for Gemini 2.5 Flash-Lite, 4.1% for Llama 3.3 70B, and 4.5% for Mistral Large 2411. The 2025 OpenAI paper Why Language Models Hallucinate says that common training and evaluation procedures can reward guessing over admitting uncertainty.
• A context window limits how much information a model can consider at once. The context window is the amount of information, measured in tokens, that a model can process for a particular interaction. If a conversation exceeds a model’s limit, older material may have to be dropped, or the request may fail. OpenAI documents both behaviors in its APIs, saying longer context can also increase latency and cost.
• The models are constrained by the data they learn from. A 2024 MIT Press Open Mind review says that LLMs’ reliance on vast quantities of written material makes them a poor analog for the full diversity of human language, particularly languages with little written data and forms such as sign language.
• A capable model requires a lot of computing power. Stanford’s 2026 AI Index notes that inference can consume more energy than the original training run when a model is used at sufficient scale. For a benchmark prompt of roughly 1,000 input and 1,000 output tokens, estimated energy consumption among select 2025 models ranged from about five watt-hours to more than 23 watt-hours per query.
• Model fluency can hide brittle performance. A 2026 paper in AI & Society classified 13 kinds of LLM failures, including incorrect inference and susceptibility to misleading prompts. “Our research shows that many of the weaknesses we see in AI today can be traced back to the fact that LLM learning approaches are not the equivalent to human socialization or participation in social life,” coauthor Simon Thorne, a senior lecturer in computer science at Cardiff Metropolitan University, told the institution.

How are large language models used in ecommerce?

In ecommerce, LLMs can be used to perform tasks like creating product copy, answering customer queries, coding, and data analysis. The common thread is language. LLMs can interpret what an online store owner or shopper is asking, generate a response, and call on other software or data sources to help complete the task. 

US Census Bureau data collected from December 2025 to May 2026 found that 17% to 20% of US businesses were using AI in business functions. The retail sector lagged the national rate—around 14% of retail businesses used AI, while about 17% expected to within six months. 

Chatbots and virtual assistants

In ecommerce, chatbots and AI assistants can interpret open-ended questions and generate natural-language responses instead of relying on a fixed menu of scripted answers. Shopify Inbox, for example, can draft AI-generated suggested replies to customer inquiries that draw on a store’s information. The Inbox agent can answer customers automatically using product catalog data, store policies, and published storefront content. For shoppers signed in with Shop, it can also personalize product recommendations using purchase history.

Content generation

In a 2025 Shopify survey, content generation was the most common AI use case among store owners at 69%.*

Shopify Sidekick can generate and edit product descriptions, blog posts, emails, marketing campaigns, and other store content from natural-language instructions. 

Children’s clothing brand Brave Little Ones uses Sidekick to turn his team’s creative ideas into storefront changes. The brand takes advantage of Shopify’s built-in AI tool to make changes to the site’s pages ahead of big sales.

“Sidekick allows us to come up with a concept and then immediately create a section and throw it on the website and see if it worked or not,” says cofounder Jon Ezell.

Personalized shopping experiences

An LLM can make product discovery more conversational. A shopper can describe their needs in ordinary language instead of filtering a catalog one attribute at a time.

A 2025 YouGov survey found that 14% of US adults had used an AI shopping assistant, rising to 24% within Gen Z. Among all users, 44% had asked product questions, 41% had used an assistant to find specific products, and 33% had sought product recommendations. 

Agentic commerce is a new way of shopping where AI assistants can carry out more of that shopping journey from discovery to purchase. Shopify’s Agentic Storefronts make your products available for purchase across AI channels, including ChatGPT, Microsoft Copilot, Google AI Mode, and Gemini.

AI referrals are a growing source of revenue. In the first quarter of 2026, referral sessions to Shopify storefronts from select AI chatbots grew eightfold year over year, while AI-referred orders rose nearly 13-fold.

Search optimization

LLMs can help ecommerce teams optimize for two kinds of search at once—traditional search engines and the growing number of AI systems that answer shopping questions directly.

That second job is called generative engine optimization, which includes improving the likelihood that products and brands appear in responses from tools such as ChatGPT, Gemini, and Google AI Mode.

“Shopify themes load fast and expose structured product data that AI systems and search engines can read, and checkout is built for shoppers who arrive ready to buy,” writes Kyle Risley, senior SEO lead at Shopify.

Shopify is also building these signals into your workflows. The Spring ’26 Edition introduced tools for showing the top AI shopping queries in your category, shows which of your products appear for them, and flags gaps. Sidekick can then recommend changes, such as adding missing specifications, improving product descriptions, or making titles more descriptive when products are being surfaced but not converting.

Outdoor furniture brand Polywood optimized their 150,000-SKU catalog for LLM indexing and joined an early ChatGPT shopping pilot. 

“You were the ones who leaned forward the most in enabling us to be ready for it,” says Benjamin Spiegel, chief digital officer of Polywood, describing Shopify’s work to optimize the company’s knowledge base for LLM indexing. 

Data analysis

LLMs can turn business data into something you can query in plain language. Furniture brand OMHU does exactly that with Shopify Sidekick. Their ecommerce team asks questions such as what is driving a weekly conversion drop, which collections are underperforming, and how new products are selling across markets. 

The workflow saved OMHU around 10 hours a month on analytics work, and the team has used Sidekick’s answers as a starting point for inventory reordering decisions. 

LLMs can also analyze large amounts of unstructured text, like product reviews, customer feedback, support conversations, and social posts, to identify how customers feel about a particular brand in what’s known as sentiment analysis. 

Automating administrative tasks

LLMs can turn plain-language instructions into actions inside business software, which makes them useful for repetitive administrative work. Sidekick can edit product pages, build Shopify Flow automations, generate reports, and create custom apps from conversational prompts. Sidekick presents changes for the merchant to review before applying them.

Apparel brand Snocks has used Sidekick for this kind of work. For the brand, building Shopify Flows went from hours of research to a single conversation, while reports that once took up to 30 minutes could be generated in seconds, resulting in up to 98% faster reporting.

“Everyone has worked with AI by now. But having it sit right inside Shopify, connected to your own data, makes it so much more usable,” says Kevin Foitzik, group head of e-commerce and IT at Snocks.

Shopify’s Smart Pricing app uses machine-learning models to recommend product markdowns and markups based on sales, inventory, seasonality, and product trends. 

Language translation

LLMs can translate product descriptions, support replies, marketing copy, and other ecommerce content while attempting to preserve more of the surrounding context than a simple word-for-word substitution.

Shopify’s Translate & Adapt app lets you translate products, collections, blog posts, pages, and other storefront content, and adapt wording for individual markets. The app can automatically translate up to two languages for free, while you can manually add additional translations and edit machine-generated copy in its side-by-side editor.

Fraud detection

LLMs can help fraud teams review cases, summarize evidence, and work through unstructured information, however, the systems that actually score transaction risk are often other machine-learning models. 

The Merchant Risk Council’s 2026 Global Ecommerce Payments & Fraud Report found that merchants lost 3.2% of annual ecommerce revenue to payment fraud globally. 

Shopify’s built-in fraud analysis uses machine-learning algorithms trained on historical transactions across Shopify stores to classify online credit-card orders as low, medium, or high risk. The native tools consider signals such as AVS (Address Verification Service) and CVV (Card Verification Value) checks, IP-address details, and attempts to use multiple cards, while Shopify Flow can automate what happens to suspicious orders.

*Based on a November 2025 survey of 500 Shopify merchants conducted in English across Australia, Canada, the United Kingdom, Ireland, New Zealand, and the United States. Respondents were established merchants with two or more years on the platform. Results reflect the experiences of this specific sample and may not be representative of all merchants.