A large language model, or LLM, is the kind of software behind ChatGPT, Claude, Gemini and most of the AI writing tools you see today. You type a request in plain language, and it answers in plain language. It can draft a product description, rewrite an email, summarize a long supplier contract or suggest ten names for a new candle scent.

This page is for sellers, not engineers. If you run a store, sell a course or manage a small brand, you probably already use an LLM, directly or through a tool that hides it. Knowing how it works helps you get better output, spend less, and avoid the errors that end up on a product page in front of paying customers.

What is an LLM?

An LLM is a statistical model trained on a very large collection of text: books, websites, code, forums and documentation. During training, it learns to predict the next piece of text given everything that came before. Those pieces are called tokens. A token is roughly three quarters of an English word, so 1,000 words is about 1,300 tokens.

"Large" refers to two things. The training data runs into trillions of tokens, and the model itself holds billions of adjustable numbers called parameters. After that first phase, called pre-training, most models go through a second phase where humans rate answers and the model is tuned to follow instructions and be helpful. That second phase turns a text predictor into an assistant.

What an LLM is not:

  • Not a database. It does not look up facts. It generates text that is statistically likely. When it lacks the answer, it can still produce a confident sentence that is wrong. This is called a hallucination.
  • Not a search engine. Unless a tool connects it to the web or to your documents, it only knows what was in its training data, which stops at a cutoff date.
  • Not a person who knows your business. It has never seen your stock levels, your return policy or your best-selling variant unless you give it that information.

Related vocabulary you will meet: the prompt is your instruction; the context window is how much text the model can read at once; fine-tuning means training a model further on your own examples. GPT is one family of LLMs, made by OpenAI. Retrieval-augmented generation is a technique that feeds the model your own documents so it answers from them.

Why it matters

For a small seller, the main resource an LLM saves is time spent writing. Writing is a large, hidden part of running a store: product pages, collection intros, email campaigns, FAQ entries, replies to the same shipping question for the fortieth time.

Take a jewelry brand with 200 products. Writing a good description by hand takes about 20 minutes per product, so the full catalog costs around 67 hours. With an LLM, you give it the material, the stone, the size, the care instructions and three examples of your tone. It drafts each description in seconds. You then spend 6 minutes checking facts and adjusting the voice. The catalog now costs 20 hours, which saves 47 hours of work.

The token bill is small. Each description needs about 1,500 tokens of input and output combined, so 200 products use around 300,000 tokens. At current API prices, that is well under $5, even on a premium model. Most sellers use a monthly chat subscription of about $20 instead and never think about tokens.

The risk is also real. A model that invents "925 sterling silver" for a plated ring creates a false product claim. That leads to returns, chargebacks and, in some countries, consumer protection issues. The time you save only counts if the review step stays in place.

How it works

Here is what happens when you ask an LLM to write a product description:

  • Your text is split into tokens. "Hand-poured soy candle, 220 g" becomes a dozen or so tokens.
  • The model reads the whole context. This includes a hidden system instruction from the tool, your prompt, any files you attached and the earlier messages in the conversation, up to the context window limit.
  • It computes a probability for every possible next token. "The" might get 12%, "This" 9%, "Our" 7%, and so on across tens of thousands of options.
  • It picks one, adds it to the text and repeats. A setting called temperature controls how adventurous the choice is. Low temperature gives predictable wording, high temperature gives more variety and more risk.
  • It stops when it predicts an end token or hits a length limit.

This is why the same prompt gives a different answer each time, and why detail in your prompt matters so much. The model cannot ask what your candle smells like. If you do not say it, it guesses.

Benchmarks and examples

Some realistic reference points for sellers:

  • Context windows range from about 100,000 to over 1,000,000 tokens on current commercial models. 100,000 tokens is roughly a 250-page book, which is more than enough for a full catalog export of a small store.
  • Speed is typically 30 to 150 words per second once the answer starts.
  • Price through an API goes from a few cents to around $15 per million output tokens, depending on the model size. Consumer chat plans sit around $20 a month.
  • Error rate on facts depends heavily on the task. When the model rewrites facts you provided, errors are rare. When it has to recall facts from memory, such as a regulation, a shipping rule or a material spec, errors are common enough that you should check every claim.

Typical uses: a creator launching a course asks an LLM to turn a 40-minute recorded lesson transcript into a sales page draft. A small skincare brand drafts answers to its 25 most frequent support questions, then edits them into saved replies. A store selling in French and English drafts a translation of each product page, then has a native speaker review the 20 best sellers.

Common mistakes

  • Publishing without reading. Invented materials, dimensions, certifications or delivery promises are the most costly errors. Every number and every claim should be checked against your real product.
  • Vague prompts. "Write a description for my hoodie" produces the same bland text as every other store. Give the fabric weight, the fit, the audience and an example of your voice.
  • Pasting private data into consumer tools. Customer names, addresses and order histories do not belong in a free chat tool unless you have checked its data policy and your privacy obligations.
  • Letting the model answer customers alone. An unsupervised bot that promises a refund your policy does not allow creates a real commitment you may have to honor.
  • Assuming it knows recent events. A model trained last year does not know your new shipping carrier or this month's promotion.

Best practices

  • Write a reusable brief. Keep one document with your brand voice, banned words, target customer and three example descriptions you love. Paste it at the start of every session.
  • Give facts, ask for wording. Supply the specs in a list and ask the model to write around them. It is much better at phrasing than at remembering.
  • Ask for several versions. Request three short variants, pick the best, and combine lines. This is faster than editing one weak draft.
  • Keep a human review step. Check numbers, claims and anything legal. Read the text out loud once before it goes live.
  • Use it for structure, not just prose. LLMs are good at turning messy notes into a clean FAQ, a size guide outline or a comparison of three suppliers.
  • Compare tools on your own task. Run the same five products through two or three models and judge the results on your catalog, not on a leaderboard.

In Roctify

Roctify does not include an AI assistant or a built-in LLM. You use the model of your choice outside the platform, then paste the reviewed text where it belongs: product descriptions, variant names, discount code announcements, email marketing campaigns and replies in the email inbox (Creator plan and up). Because Roctify keeps one shared catalog, a description you improve once is updated on your storefront and your link-in-bio page at the same time.

If you sell in several languages, Roctify's multi-language settings let you store each version of a product text. An LLM can draft those versions, and you or a native speaker can review them before you publish. Pro plan reports and exports also give you a clean list of products and sales, which is useful material to hand to a model when you ask it which descriptions to rewrite first.

FAQ

Can I use AI-written product descriptions without hurting SEO?

Yes, as long as the text is accurate, specific and useful to the buyer. Search engines penalize thin, repetitive or misleading pages, not the tool used to write them. A description that states real dimensions, materials and use cases in your own voice performs better than generic filler, whoever typed it.

Which LLM should a small store choose?

Start with the chat tool you already pay for, and test it on five of your real products. Judge the result on accuracy, tone and how much editing it needs. Switching models is easy, so there is no reason to commit early or pay for several at once.

Is it safe to paste customer messages into an LLM?

It depends on the tool and your settings. Many business plans do not use your data for training, while some free plans may. Remove names, emails and addresses before pasting, and check what your privacy policy promises your customers.