Measurement
AI search visibility metrics that support decisions
A useful GEO metric does not just look precise. It tells a team what to investigate next, and it remains honest about the sampling limits behind the number.
AI search reporting is full of confident-looking dashboards. The risk is false precision: one prompt run becomes a “score,” one screenshot becomes a trend, and an unknown source becomes a zero. A better measurement system starts with the decision the metric should support.
Six core metrics
| Metric | Definition | Decision it supports |
|---|---|---|
| Mention rate | Share of sampled answers that name the brand | Is the brand visible at all in this prompt set? |
| Recommendation position | Average position when the brand appears in a shortlist | Is it a primary option or a late mention? |
| Source coverage | Share of answers with a traceable supporting source | Where does the brand need stronger public evidence? |
| Entity consistency | Agreement across answers on core brand facts | Are conflicting descriptions weakening the entity? |
| Content coverage | Coverage of the prompt set by owned, useful pages | Which questions lack a clear brand answer? |
| Change over time | Movement under a fixed sampling frame | Did the intervention produce a durable shift? |
Sampling is part of the metric
Generative answers are not deterministic. The same prompt can return different brands, different wording, and different sources. A metric without a sampling protocol is difficult to defend.
- Fix the prompt set. Keep the wording stable across comparison periods.
- Repeat each prompt. Multiple runs distinguish a stable pattern from a one-off answer.
- Record engine and region. Model versions and regional availability change results.
- Store the raw answer and sources. An aggregate should be traceable to evidence.
Report a range or a stability indicator when the sample is small. “We saw this twice out of three runs” is more useful than a single decimal that implies certainty.
Why recommendation position is not enough
A brand can be mentioned in every answer and still be a weak recommendation. Position matters, but so does the sentence around it: “a niche option,” “a cheaper alternative,” and “the leading platform” are not equivalent. A GEO report should keep the context, not just the rank.
Source coverage and citation quality
An answer may cite the brand’s own website, a review site, a directory, or a third-party article. Source coverage asks whether a claim can be checked. Citation quality asks whether the source is appropriate for the claim.
Do not treat every link as equal. A product page can support a feature claim; an unrelated directory may not. When the answer has no explicit source, mark it as an evidence gap rather than inventing an attribution.
Five common measurement mistakes
- One prompt, one conclusion. A single answer is an anecdote.
- Changing the prompt set. A “gain” may simply be an easier question.
- Using rank without context. Position without recommendation language is incomplete.
- Treating missing data as zero. Unknown evidence should be excluded or flagged.
- Confusing model wording with brand quality. An AI answer is an observation, not an objective audit of the company.
A one-page measurement plan
- Define 10–30 decision prompts for the category.
- Choose the engines that matter to the market.
- Run the same panel at a fixed cadence.
- Store prompts, answers, sources, engine, region, and timestamp.
- Review mention, position, source coverage, consistency, and change together.
- Publish one intervention at a time when possible, then remeasure.
Glaruno implements this panel in its diagnostic and monitoring workflow. If you need a starting point, read what GEO is or compare the early-access plans.
Frequently asked questions
What is a good mention rate?
There is no universal benchmark. It depends on the category, prompt set, engines, competitors, and sampling depth. The useful comparison is the same brand and prompt set over time.
Why repeat the same prompt multiple times?
Generative outputs vary. Repeated samples reveal how stable an answer is and prevent a single unusual response from being treated as a trend.
Should I optimize for every AI engine?
Start with the engines that matter to the target market and are supported by the deployment. Adding engines increases coverage but also changes the measurement frame.