Grounding answers built from live sources.

Grounding is when an AI system builds its answer from sources it retrieves at the moment you ask, rather than from the model’s trained-in memory. A grounded answer can cite specific pages and is far less likely to invent facts; an ungrounded answer is the model recalling what it absorbed during training, with no live source behind it. Prefer tracks which pages the engines cite for your buyer questions.

How grounding works#

When a prompt arrives, a grounded engine searches the web (and any structured data it has), pulls the passages that best answer the question, and writes its response from that retrieved material, citing the sources. This is the same retrieval step behind query fan-out, which Google describes in its documentation on AI features. An ungrounded answer skips retrieval and generates from memory alone.

A 4-step path. 1. Retrieve (The engine searches the web and its data for sources that match the question); 2. Read (It pulls the passages that answer the question most directly); 3. Synthesize (It writes the answer from the retrieved material, not from memory); 4. Cite (It names the sources it used, which is where your page can appear).

Most answers to real buyer questions are grounded today. In our first GPT-5.6-era measurement run (50 prompts, 2 September 2026), 47 of 50 answers were grounded in retrieved sources; only 3 were answered from memory.

Bar chart. Web-grounded 47answers, From memory 3answers.

Why grounding matters for AI visibility#

You can only be cited in a grounded answer, because citation requires a retrieved source. When an engine answers from memory, there are no sources to be one of, and whatever it knows about you is frozen at training time and may be wrong. When it grounds, your page can be the source it quotes.

The three ungrounded answers in our run share a trait: they were short definitional prompts the model felt no need to look up. For those, the win condition is different. You need the model to hold accurate facts about you in memory, which comes from being a well-formed knowledge graph entity. We saw both failure modes on one prompt over time: “What is Prefer?” was answered by fabrication on our day-zero baseline, then by a grounded, cited answer from our own domain in this run.

Two columns. An answer from memory, generated without retrieval: Recalls what the model absorbed in training, Has no sources to cite, Can be stale or invented, Fixed only by changing the model's memory. A grounded answer, built from retrieved sources: Uses pages found at query time, Cites the sources it quotes, Reflects current, checkable facts, Fixed by improving the page it retrieves.

Grounding vs hallucination#

A hallucination is a confident, wrong answer with nothing behind it, the classic failure of an ungrounded model. Grounding is the main defense: by forcing the answer to come from retrieved sources, the engine has something real to quote and something to cite. Grounding does not eliminate errors, but it ties the answer to a checkable source.

Example#

Ask an assistant “what does [your product] cost?” A grounded engine retrieves your pricing page and quotes the current figure with a citation. An ungrounded engine recites whatever price it saw in training, which may be a year stale. The first can be corrected by fixing your page; the second can only be corrected by becoming a strong enough entity that the model’s memory is right.

How to earn grounding#

Summary graphic of 4 items: 1. Publish extractable passages: Lead each section with a self-contained answer a model can lift and quote directly. 2. Keep facts current and dated: State prices, specs, and numbers plainly and keep them fresh, so the retrieved passage is right. 3. Be a clear entity: Consistent name, category, and facts so the model recognizes and trusts what it retrieves. 4. Earn trusted sources: Get corroboration on the third-party pages engines already retrieve for your category.

How Prefer helps#

Prefer tracks whether each engine grounds its answer about you and which source it quotes, and ships the extractable content and entity work that make your page the one retrieved. Run a free AI visibility audit to see how AI engines answer about you today.