Grounding

Grounding is the practice of tying a model's answer to retrieved source material, so that what it says can be traced to something outside its own weights. It is the mechanism behind every AI answer that arrives with links attached.

Google's Gemini API documentation describes its own implementation in terms of three benefits: increasing factual accuracy by basing responses on real-world information, accessing real-time information so recent events can be answered, and providing citations so users can see the sources for the model's claims. It notes that grounding with Google Search connects the model to real-time web content, works with all available languages, and lets it cite verifiable sources beyond its knowledge cutoff.

The loop, as documented

Enabling a search tool hands the whole retrieval workflow to the model rather than to the application. The documented sequence is:

1. The application sends a user prompt with the search tool enabled.
2. The model analyses the prompt and decides whether a search would improve the answer.
3. If so, it generates one or several search queries itself and executes them.
4. It processes the results, synthesises the information, and formulates a response.
5. It returns a response grounded in those results, including inline annotations carrying the citations, together with the search call and search result steps showing the queries it ran.

Two consequences follow from step 2 and step 3 that are easy to miss. The model decides whether to search — some prompts are answered from parameters alone — and the model writes the queries. Nobody is typing your brand name into a box; a machine is inventing the phrasing it thinks will find the answer, and your page is competing on how it reads against those invented queries.

Grounded is not the same as verified

The word invites more confidence than it earns. A citation records which document was retrieved and associated with a span of output. It does not certify that the document supports the sentence, that the document is correct, or that the model read it the way a person would.

That is the same split described under retrieval-augmented generation: retrieval can return a passage that is topically close but answers a different question, and the model will then write confidently from it. Google's own framing is careful about this — the claim is that grounding reduces hallucinations, not that it removes them.

For anyone auditing what assistants say about their company, this distinction decides where the fix goes. If the citation points at a page of yours that genuinely says the wrong thing, edit the page. If it points at a page that says the right thing and the answer says otherwise, no amount of editing will help, and the useful record is the transcript.

Why it is the important mode for publishers

An ungrounded answer is written from training data frozen at a cutoff. Nothing published since can reach it, and any URL in it is reconstructed from memory rather than fetched — which is how confident-looking dead links appear.

A grounded answer changes all three properties at once. Your current page is a candidate, the answer can carry a real link, and an error you correct is corrected in the next answer rather than the next model. That is the entire practical case for treating AI answers as something you can influence rather than merely observe.

It also puts access upstream of everything else. Grounding begins with a fetch, so a page an agent may not fetch is not in the running however well it is written. What robots.txt says, whether the origin is fast, and whether the content survives without JavaScript all decide eligibility before quality is considered.

Writing for it

The unit that gets grounded is a passage, not a page, so the practical advice is the same shape as for retrieval generally:

• Give each section one subject and a heading that states its claim, so it survives being lifted out.
• Put the qualifier in the same sentence as the claim; a caveat two paragraphs up does not travel.
• Keep names, version numbers and identifiers literal, because a query the model invented may contain them verbatim.
• State the date on anything time-sensitive, since a grounded answer will happily present a stale fact as current if the page never says when it was true.

The last one is worth taking seriously. Grounding attaches your page to an answer; it does not attach your caveats.