Grounded Response Rate
- Grounded Response Rate = share of answers backed by retrieved sources, not model memory.
- It tells you which lever works: content/citations for grounded topics, anchoring for parametric ones.
- High grounded rate means faster to win and faster to lose; parametric topics are sticky both ways.
- It is a per-topic, per-provider diagnostic with a known blind spot — silent retrieval.
Consensus definition
AI answers are produced on a spectrum from purely parametric (drawn from training weights) to fully grounded (conditioned on documents retrieved at answer time).1 Grounded Response Rate measures where a topic sits on that spectrum: across a prompt set, the proportion of answers that actually retrieve and cite live sources. It matters because the two regimes respond to different work. Where answers are grounded, fresh and well-structured content can change them; where answers are parametric, only deeper anchoring over time can. Measuring the rate tells a brand which lever is even available.
rhinegold operator refinement
Rhinegold treats Grounded Response Rate as a diagnostic that should precede GEO tactics. A common failure is to pour effort into content optimisation for a topic the models answer from memory — the documents are never retrieved, so the work cannot land. Conversely, a high grounded rate is an opening: it means retrieval is active and current content has a path into the answer. The rate is also a stability signal. Heavily grounded answers move with the source landscape — faster to win, faster to lose; parametric answers are sticky in both directions. Reading Citation Rate without the grounded rate behind it confuses 'the model cited a source' with 'this topic is decided by sources at all.'
Operational use
Estimate the grounded rate per topic cluster by inspecting whether answers carry live citations or read as memory-based, across prompt variants and providers. Use it to triage: grounded clusters are addressable with content and citation work now; parametric clusters are an anchoring programme measured in quarters.
Measurement boundary
Grounding is partially observable — a cited source signals retrieval, but its absence does not prove parametric generation, since some systems retrieve without surfacing citations. The rate is therefore an estimate with a known blind spot, is provider-specific (Platform Divergence), and says nothing about whether the grounded source was the brand's own.
Distinct from
Against Grounding, which is the retrieval mechanism for a single answer — the grounded rate aggregates how often that mechanism fires across a topic. Against Citation Rate, which measures whether the brand is the cited source; the grounded rate measures whether any source backs the answer. Against Semantic Anchoring, its conceptual inverse: a low grounded rate is precisely the regime where anchoring, not retrieval, governs visibility.
Observed pattern in practice
Common mistakes
- Optimising content for a parametric topic — the documents are never retrieved, so the work cannot reach the answer.
- Reading citation presence as proof of grounding everywhere — some systems retrieve silently, so the rate has a blind spot.
- Averaging the grounded rate across topics and providers — it is a per-cluster, per-surface diagnostic, not a single site number.
Where consensus is missing
There is no standard method to classify an answer as grounded versus parametric from the outside, no agreed threshold separating the regimes, and no cross-provider benchmark. The metric is an emerging practitioner construct, not a settled KPI.
Sources & deeper reading
- rhinegold grounded-rate triage methodology — per-cluster regime classification to route GEO effort, shared with engaged clients and partners.