Three places an answer can come from
Training: what it learned before you arrived
Every model has a training cut-off: a date after which it has read nothing. Ask about anything newer and a plain model can only guess. Training also reflects what was in the data. If writing from some countries, languages or eras is over-represented, answers lean that way. Ask for "great books to read" ten times and the lists may keep favouring one country and one century. That is a pattern in the data, not the model's taste.
Search: reading the web for you
Many chatbots can now run a web search and read pages before answering. That helps with recent events and prices, but it adds a new question: which pages did it read? A good product shows its sources as links. A confident answer with no links tells you nothing about whether it searched at all, so look for the evidence on screen rather than judging by how up to date the wording sounds.
Your files: reading what you give it
When you upload a document or paste text, the model answers from that material, which is usually more reliable for questions about that document. It can still misread tables, skip pages or blend in outside knowledge. Ask it to quote the exact passage it relied on, then check the quote is really there. Scanned PDFs are a common trap: if you cannot select the text yourself, the tool may not be able to read it either.
Model, product and provider
The model is the trained engine. The product is everything wrapped around it: the chat window, hidden instructions that set its tone, memory, search and file tools. The same or a similar model can sit behind very different products, which is why two apps can behave differently. Some models are only available through a company's service; others are openly published so anyone can download and run them, even on a laptop.
Think of the last answer an AI gave you. Do you know whether it searched the web, read your file, or answered from training alone?