Grounding

Updated 2026-09-15

Grounding is the practice of generating an AI answer from documents retrieved at query time instead of from the model's trained-in memory. A grounded answer can cite what it used; an ungrounded one is reciting whatever the model absorbed during training.

How it is calculated

Grounding is not a rate but a status assigned to each sampled answer before any metric is computed. An answer that triggered a search and returned sources is grounded; an answer produced without retrieval is a memory answer; a surface that did not appear at all is neither. Only grounded answers belong in visibility rates, because only they reflect what the engine can currently find about you.

Worked example (illustrative statuses)

Ten samples of one prompt return: 7 grounded answers, 2 memory answers and 1 provider error. The mention rate for that prompt is computed over the 7 grounded answers. Folding the 2 memory answers in would measure the model's training data, and counting the error as a miss would measure the API's uptime.

Common mistakes

  • Treating a memory answer as evidence of visibility — it reflects a training snapshot, not today's web.
  • Assuming every engine grounds by default. Some always ground, some decide per query, and the decision changes what your number means.
  • Ignoring the grounded share over time: a falling grounded ratio usually means the engine changed behaviour, not that your visibility moved.

Frequently asked questions

Which engines ground their answers?
Behaviour differs and changes: some models always retrieve before answering, others retrieve only when the query looks like it needs fresh information, and AI Overviews are generated inside the search result page itself. Any serious measurement records the grounding status per answer rather than assuming one behaviour.
Why exclude memory answers from visibility metrics?
Because they answer a different question. A memory answer tells you what the model learned months ago; a grounded answer tells you what the engine can find and cite about you right now, which is the thing content work can actually move.

Related terms