What is Grounding?
Grounding is when an AI model backs its answer with live retrieved sources — running a web search mid-answer and reading real pages — instead of answering purely from what it memorized in training.
How often do AI answers actually ground?
Far less than assumed. Models ground when they judge a question to need current information; for evergreen category questions — 'best X for Y' — they frequently answer from memory. In SeeGeo's own measurement, a model offered live search used it on 1 of 20 category prompts, answering the other 19 from training data alone. An ungrounded answer carries no citations, so 'not cited' often really means 'nothing was looked up'.
Why does grounding matter for visibility?
Because it splits visibility into two games. Ungrounded answers draw on training data — you win those by being well-documented on the web over time, which is slow. Grounded answers draw on a handful of freshly retrieved pages — you win those by being present on the exact pages retrieval selects, mostly third-party comparisons and directories. A serious visibility strategy plays both.
Can you make an answer ground?
Sometimes — questions that demand current information ('pricing in 2026', 'latest') trigger search far more often than evergreen ones. But you don't control how customers phrase things, so treating grounding as a switch you can flip is a mistake. The robust posture plays both games: be well-documented enough to win ungrounded answers from memory, and present on retrieval-trusted pages for the grounded ones.
Frequently asked questions
What does Grounding mean?
Grounding is when an AI model backs its answer with live retrieved sources — running a web search mid-answer and reading real pages — instead of answering purely from what it memorized in training.
How can I tell if an answer was grounded?
Citations are the tell: grounded answers can link the pages they read, ungrounded ones have nothing to link. Perplexity grounds nearly always; ChatGPT and Gemini decide per question, which is why the same prompt sometimes returns sources and sometimes none.
Does grounding use my robots.txt rules?
Yes — grounded retrieval reaches your site through the engine's crawlers and fetchers (OAI-SearchBot, ChatGPT-User, Google's infrastructure), all of which honor crawler rules. Blocking them blocks grounded answers from ever reading you.
Why did my citation disappear between two identical questions?
Most likely the first answer grounded and the second didn't — the model chose memory the second time, and an answer from memory has no citations to give. This is why honest citation tracking distinguishes 'not cited' from 'nothing was retrieved at all'.