Grounding

What does a grounded answer actually carry?

More than the visible citation list, and Google’s Gemini API is the one place you can see it. A grounded response returns the search queries the model ran in a google_search_call block, and annotates its own text with url_citation entries carrying the source URL, the title, and start_index and end_index marking exactly which stretch of the answer that source supports.

That last detail is the useful one. Grounding is not attached to the answer as a whole, it is attached to spans of it, so a source is credited for a specific claim. It is the closest thing to a receipt that any of these systems produces. It also means a page can be cited for one sentence in a long answer and contribute nothing else, which is what most citations actually are. Reading a grounded response as an endorsement of the whole answer overstates what the markup claims.

How much of a page does grounding actually use?

Less than most people write, and the largest measurement of it is Ahrefs’. Analysing 560,346 AI Overviews and the 1,677,876 URLs they cited, it found the average cited page runs to 1,282 words, that 53.4% of cited pages are under 1,000 words and only 16% are over 2,000, and that the correlation between word count and citation position is 0.04, which is nothing at all.

The consequence is the one this glossary keeps arriving at from different directions. Length is not the lever, and a long page is not better grounded, it is more thinly used: grounding attaches a source to a span of the answer, so what it needs from you is one passage that settles one question. A 3,000-word page rarely holds more of those than a 900-word one, it just holds more words between them.

Why is grounding not the same as being right?

Because grounding constrains where an answer came from, not whether the source was correct. A model can ground a claim perfectly in a page that is out of date, mistaken or somebody’s marketing copy, and the citation will look exactly as trustworthy as a good one. It reduces invention, which is real and valuable, and it does nothing about error.

For a publisher that cuts a specific way: a wrong page of yours is more dangerous once grounding exists, not less, because it now gets quoted with your name on it.

Related terms

• RAG (Retrieval-Augmented Generation)RAG, or retrieval-augmented generation, is the technique of fetching documents at question time and giving them to a language model to answer from, rather than relying on what the model learned in training.
• ChunkingChunking is the step that splits a page into passages so a retrieval system can select one without the rest.
• Query Fan-OutQuery fan-out is Google’s own term for how AI Overviews and AI Mode answer a question: the engine breaks the question into subtopics and issues many searches at once, then writes one answer from the results.
• AI CitationAn AI citation is a link an AI assistant attaches to its answer, naming a page as a source.