58%
What it is
A large language model is a program trained on a huge amount of text: websites, books, articles and more. From that text it learns how words fit together. When you ask it something, it writes an answer one word at a time, picking words that are likely to come next.
ChatGPT uses OpenAI's GPT models. Gemini and Google's AI search features use Google's Gemini models. Claude is made by Anthropic.
Where its answers come from
An LLM can answer from two places:
- What it learned in training. This is fixed at a point in time, called the knowledge cutoff. Big, well-known brands are often in it.
- Pages it reads now. AI search tools search the web and give the model a few pages to read before it answers. This is retrieval-augmented generation.
Most product recommendations use both. In our study, ChatGPT put brand names into its web search before reading anything for 109 of 210 questions. It already had a shortlist from training, then used the pages it found to fill in the rest.
Why answers change
LLMs don't give the same answer every time. They pick words with some randomness, and AI search tools read different pages on different runs. In our study, 84% of answers named a different set of brands the next day. That's why one check of "does ChatGPT recommend us?" means little.
Example
Ask an LLM "what's a good CRM?" with search turned off, and it answers from what it learned in training: brands that came up often in the text it was trained on.
Ask the same question with search on, and it first reads today's pages, then writes. The second answer can include newer brands and current prices. That's the difference between training data and RAG.
Frequently asked questions
Can I get my brand into an LLM's training data?
Not directly. Models learn from what's widely written on the web, so being covered on many sites helps over time. But that's slow, and you can't check when it happens. Getting onto the pages AI reads now works faster and can be measured.
Do LLMs know my prices and features?
Only if they've read them, either in training or on a page they find during the search. If your prices aren't on a page they can read, the AI may guess or leave you out.
Which LLMs matter for AI search?
The ones behind the tools buyers use: OpenAI's GPT models in ChatGPT, and Google's Gemini models in Gemini, AI Overviews and AI Mode. Claude and others power smaller shares.
Why does the same question get different answers?
LLMs pick each word with some randomness, and search results change. Together that makes answers shift from one run to the next. See AI answer volatility.