Memory answer
Updated 2026-09-15
A memory answer is one the model produces without retrieving anything, relying on what it absorbed during training. It reads exactly like a grounded answer and means something entirely different for measurement.
Why it is recorded separately
A memory answer describes the web as it was when the model was trained, so it lags reality by months and cannot reflect anything you published recently. Mixing memory answers into a visibility rate produces a number that partly measures your content and partly measures a historical snapshot — and the mixture ratio changes as engines adjust when they search. The status is therefore recorded per answer, and visibility rates are computed over grounded answers only.
Illustrative split (example numbers)
Thirty samples of one prompt: 23 grounded, 7 memory. The brand appears in 4 grounded and 5 memory answers. Reported over grounded answers the rate is 4/23; reported over everything it is 9/30 — a materially better-looking number describing something else.
Common mistakes
- Celebrating high visibility that comes mostly from memory answers describing an old version of the product.
- Ignoring memory answers entirely — they are where stale or wrong descriptions surface, which is worth reading even if not counted.
- Assuming the grounded share is stable across cycles; it drifts with engine behaviour.
Frequently asked questions
- Can we make an engine search instead of answering from memory?
- Not from the outside. Phrasing that implies recency makes retrieval more likely, but the decision belongs to the engine — which is why the status is measured rather than assumed.
- Are memory answers useless?
- No: they show how the model learned to describe your category and where stale claims persist. They are simply not evidence about your current visibility.
