How grounding works
A language model on its own writes text that is statistically likely, not text that is checked. Grounding changes the order of operations. Before answering, the system searches something real — the live web, a search index, a company’s own documents — pulls back the passages that look relevant, and composes the reply from those passages rather than from memory. The links or footnotes under an AI answer are the visible end of that process.
The retrieval step is where the outcome is decided. If a passage from your site is fetched and handed to the model, your wording can end up inside the answer and your page can end up in the source list. If nothing of yours is fetched, the model falls back on whatever it absorbed during training, which may be stale, second-hand, or actually about a competitor. This is the design behind retrieval-augmented generation, and behind the citation lists in AI search results.
Why grounding matters
It decides whether your business gets described from your own pages or from somebody else’s summary of you. A grounded system quotes a source it can point at; an ungrounded one improvises confidently. For a business that has changed its services, moved premises or dropped a product line, that is the difference between an enquiry arriving with the right expectation and one arriving with the wrong one.
It also reframes what you are optimising for. The old question was whether a page could rank. The new one is whether a specific sentence on that page is fetchable and usable when an assistant needs it, which is the thread running through most practical AI search optimisation work.
Where grounding goes wrong
The commonest failure is a confident answer grounded in the wrong source: an outdated directory listing, a scraped copy of your site, a forum thread from years ago. The citation makes it look verified, and readers rarely open it. The second failure is partial grounding — the system retrieves one paragraph, misses the qualification further down, and states something in a form you would never have used.
On the publishing side, the mistake is thinking grounding is a setting. You cannot switch it on for Google or ChatGPT, and no plugin grants it. What you control is narrower and more useful: whether the passage a system would need exists, on a page it can reach, in language it can lift without editing.
How to act on it
Write the facts you want repeated as standalone sentences that survive being pulled out of context — what you do, where you work, who you serve, what happens when someone gets in touch. Keep those details identical across your website, your Google Business Profile and any listing that carries your name, because a grounded system will find the contradiction and simply pick one version.
Then check rather than assume. Ask the assistants your customers actually use a question they would actually type, read the answer, and open every source it cites. If the sources are yours and the facts hold, the plumbing works. If an old listing or a competitor’s page is being quoted instead, you have found a specific page to write or a specific record to correct, which beats a vague worry about AI.