What is Grounded vs ungrounded?

Ask two AIs the same question and you can get two kinds of reply. A grounded answer is built by searching the live web, reading pages, and often citing them, Perplexity does this by default, and search-enabled modes of ChatGPT and Gemini do too. An ungrounded answer comes from the model's training alone: fluent, but frozen at whatever it learned, with no fresh reading and no links.

The distinction is central to GEO for two reasons. First, citations only appear in grounded answers, the model can only link a source it actually fetched. Second, grounded answers are where new or updated content can influence the result quickly, because the engine is reading the web right now rather than reciting memory.

That shapes how you measure and what you build. You measure GEO on grounded engines, because an ungrounded answer reflects old training, not your current presence, treating "the model didn't mention me" from an offline answer as evidence would be misleading. And you build content a searching engine can find, read, and quote today, knowing that's the surface where your work actually moves the needle.