Grounding
Grounding is when an AI platform bases its answer on sources it reads at the moment you ask, like web pages, instead of only on what it remembers.
Grounding is when an AI platform bases its answer on sources it reads at the moment someone asks, usually pages from a web search, instead of answering only from what it learned in training. Google's crawler documentation describes it as providing content from its search index to the model at the time of the prompt.
For a brand, grounding is the reason today's web matters to AI answers. A grounded answer can pick up your new prices, your latest reviews and last week's news. An answer from memory can't.
Grounded and ungrounded answers
An ungrounded answer comes from the model's memory. It can be fluent and confident, but it may be out of date, and if it's wrong there's no page to trace it back to.
A grounded answer comes from what the platform just read. It often links to those pages, so you can check where each claim came from. It's only as good as the pages it read, though. If those pages are wrong about you, the answer will be too.
An example
For example, someone asks "Does Fernhill Coffee deliver to Ireland?"
Answering from memory, a platform might say no, because that was true when its training data was collected. Grounded in a search, it reads Fernhill's current delivery page, sees that Ireland was added in the spring, says yes and links the page.
How platforms decide
Some platforms ground every answer. Others decide for each question whether a search is worth it. A search helps most when the answer depends on current facts, like prices, opening hours or recent news.
Grounding in AnswerOrbit
How we count shows which platforms always search the web and which decide for themselves. When an answer was grounded, you'll see the pages it used in Answers and across your prompts in Sources. When an answer said something wrong about you without using any page, Inaccuracy drivers in Claims lists the platform's own knowledge as the source.