Large language modelLLM

How do large language models work?#

A large language model works by predicting the next token — a word or word fragment — over and over, using patterns learned from billions of pages of text, then fine-tuned with human feedback so its output reads as a helpful answer.

Modern LLMs are built on the transformer, an architecture introduced by Google researchers in 2017 (Vaswani et al.) that lets the model weigh every word in a passage against every other. Scale did the rest: OpenAI’s GPT-3 paper described a model with 175 billion parameters trained mostly on filtered web crawl text (Brown et al., 2020). A second stage — fine-tuning on human preferences (Ouyang et al., 2022) — turned raw text predictors into assistants that follow instructions.

• Pre-training: the model reads a huge corpus — see LLM training data — and learns which words tend to follow which. What it absorbs becomes its built-in knowledge.
• Fine-tuning: people rank the model’s outputs and it is tuned toward the answers they prefer. This shapes tone, format and caution more than facts.
• Generation: at answer time the model writes one token at a time. It isn’t looking anything up unless a search tool hands it sources, which is why it can state a wrong price with total confidence — an AI hallucination.

Do large language models search the web?#

Large language models search the web only when the product around them decides a question needs current or specific information; stable questions are answered from training alone, and independent tests found ChatGPT and Claude searched on roughly a third of prompts.

Search is a separate tool the model can call. Anthropic’s web search documentation spells out Claude’s rule: it searches for recent events, current prices and “information about specific organizations, people, or products that might have changed,” and answers directly for established facts, creative writing and analysis. OpenAI says ChatGPT “will choose to search the web based on what you ask.” When the model does search, it follows the retrieval-augmented generation pattern: fetch sources, then write from them.

Read those rules from a buyer’s side. Questions about a specific product — its price, its integrations, how it compares — are exactly the kind the vendors say trigger a search, so they’re answered from live pages. Broad definitional questions are the ones most likely to be answered from memory.

LLM vs search engine: what's the difference?#

A search engine retrieves and ranks existing pages, while a large language model generates new text — and AI search products combine the two, using a search index to find sources and an LLM to write one answer that cites a few of them.

The practical consequence: an AI search engine is only as good as what its retrieval step finds, and each one retrieves from a different place. Perplexity runs its own 200-billion-URL index, Claude searches Brave, ChatGPT draws on Bing and OpenAI’s own index, and Google’s AI Overviews use Google’s index. The model writes the answer; the index decides who is eligible to be in it. The engine-by-engine guides cover each source.

How do you get a large language model to mention your brand?#

Get a large language model to mention your brand by working on both of its memories: make your pages easy to retrieve and quote for the questions that trigger a search, and build a consistent presence across the web so future models learn your name in training.

1. Allow the search crawlers. OAI-SearchBot, Claude-SearchBot, PerplexityBot and Googlebot feed the indexes models retrieve from. Blocking training bots like GPTBot is a separate decision. See AI crawlers.
2. Serve the text in the HTML. OpenAI’s, Anthropic’s and Perplexity’s fetchers don’t run JavaScript, so server-side rendering is the price of entry.
3. Answer the buyer’s exact question early. Engines lift passages, not pages. Write answer-first sections that name the product and state the fact.
4. Be talked about elsewhere. Reviews, forums, press and video are what both retrieval and future training see. See brand mentions.
5. Check the result. Run a fixed set of buyer prompts every week and record who gets named — that’s prompt tracking.

Related terms#

• RAGHow LLMs answerRetrieval-augmented generationA technique in which an AI system first retrieves relevant documents from an index and then gives them to a large language model to write its answer, so the response can cite current sources instead of relying only on what the model memorized in training.Read the entry
• How LLMs answerKnowledge cutoffThe date after which a large language model has no information from its training data, so anything newer — a launch, a price change, a new competitor — reaches its answers only if the model retrieves it from the live web.Read the entry
• How LLMs answerLLM training dataThe body of text — web crawls, books, code, licensed and forum content — that a language model learns from before release, and it shapes what the model says about a brand without searching, in a way that can’t be edited until the next model is trained.Read the entry
• How LLMs answerAI hallucinationA confident but false statement produced by a large language model — an invented statistic, feature, price or citation — which happens because the model generates plausible text rather than looking facts up.Read the entry
• How LLMs answerGroundingThe process of tying an AI model’s answer to specific, verifiable sources — usually web pages retrieved at the moment of the question — so the claims in the response can be checked and cited, rather than generated from the model’s memory alone.Read the entry
• AI search & GEOAI search engineA search product that answers a query with a response written by a large language model, grounded in web pages it retrieves and usually citing them, instead of returning a ranked list of links for the user to open.Read the entry