Every AI answer is built from something. Understanding what that is explains most of what looks random about AI visibility. This is general guidance; each platform decides for itself how it answers.
Two ways a model knows about you
Training data (memory): what the model absorbed when it was trained. It's frozen at a cutoff date, has no citations and can be out of date. There's no page you can edit to change it directly.
Retrieval (live search): what the assistant fetches at question time, such as search results or an index. This reads pages as they are today, and it's where most practical optimization work happens.
Most answers blend the two, which is why a brand can be named confidently but described with outdated details.
How the main engines differ
ChatGPT: with search enabled, it retrieves live results and cites sources inline. Without it, answers come from memory with no citations.
Perplexity: retrieval-first. It searches, reads a handful of pages and cites them densely. Precise pages that fully answer a specific question do well.
Google AI Overviews: grounded in Google's index and usually cite pages that already rank for the query. Conventional SEO strength matters here.
Gemini: grounded in Google Search results with the model's own knowledge layered on top.
The kinds of sources that appear most
Community discussion: Reddit, forums, Stack Overflow
Review and directory sites: G2, Capterra, Trustpilot and industry equivalents
Reference sites: Wikipedia and similar
Documentation: official docs for how-to and capability questions
Independent editorial: trade publications and analyst-style write-ups
Your own website: usually cited for facts such as pricing and features
A useful rule: assistants cite you for facts about you, and cite others for judgments about you. You can't declare yourself the best option; third-party sources carry that weight.
What this means in practice
Make sure the right pages exist and clearly answer the questions buyers ask.
Keep your brand facts consistent everywhere.
Earn presence in the third-party sources your category is answered from.
Use Citera's Sources views to see exactly which domains AI answers in your category rely on.