# How an Assistant Knows About Your Brand — AI Visibility and LLM Brand Discovery

Source: https://www.geekswithgeeks.com/en/ai-visibility/vis-how-llms-know

> Distinguish knowledge learned in training from information retrieved live from the web.

## Two sources of knowledge

A language model learns patterns from a huge training dataset up to a **cutoff date**; what it says about your brand from memory reflects how often and how consistently you appeared in that data. Many assistants can also **search the web or your own documents** at question time (retrieval, as in RAG), then write an answer from the pages they fetched, often citing them. So visibility has two parts: being **well represented in the data the model learned from**, and being **easy to find and use when it searches**. Changes you make today can show up quickly in retrieval, but only slowly (or never) in a model's built-in memory.

## Which lever affects which source

Treat this as a simplified model. Real systems mix both and change often.

```text
Source           How it works                      What you can influence
Training memory  learned before a cutoff date            long-term presence: widely cited, consistent facts
Live retrieval   search + fetch pages at question  crawlability, clear pages, fresh content, good titles
User-provided    pasted docs / connected tools      accurate public docs and help pages people paste
```

## Ask the assistant what it knows

Try asking an assistant with web search off, then on. The difference shows how much your visibility depends on retrieval versus memory, and what it currently gets wrong about you.

**Quiz:** Which source of knowledge can your recent website changes influence fastest?

- [x] Live retrieval at question time
- [ ] Training data already frozen at a past cutoff
- [ ] The model's weights directly
- [ ] Neither can be influenced

*Answer:* Live retrieval at question time. Retrieval reads the current web, whereas memory reflects older training data.