Key facts
- Answers draw on two things: what the model learned in training and pages retrieved live from the web.
- Assistants often run several related searches behind one question, sometimes called query fan out.
- Only some of the pages an assistant reads are shown as citations in the answer.
- The same question can get different answers from one ask to the next, and from one engine to another.
From question to answer
- 1The assistant reads the question and decides whether it needs fresh information from the web.
- 2If it does, it writes one or more search queries, often more specific than the question itself.
- 3It retrieves pages from a search index and reads the most relevant passages.
- 4It writes an answer that combines those passages with what the model already knows.
- 5It names brands, products or sources, and may link to some of the pages it used.
How the main engines differ
- ChatGPT can answer from training or search the web, and shows sources when it does.
- Google AI Overviews and AI Mode build answers on top of Google Search results.
- Perplexity searches for almost every question and lists its sources prominently.
- Gemini, Claude, Copilot and Grok each have their own retrieval and their own way of citing.
Because each engine retrieves and weighs sources differently, a brand can lead in one and be missing in another. That is why tracking has to be per engine.
What this means for brands
To be named, your brand needs to appear in the passages the assistant reads for that question. That comes from your own pages answering the question clearly, and from other sites the assistant trusts describing you accurately. Being crawlable, specific and consistent is the foundation.
Put this into practice with Variask Prompt research and Variask AI visibility. See Variask pricing or read the Variask FAQ.