AI visibility is observable in the answers a defined set of buyer questions receives. The useful baseline records what was asked, what came back, and how it was counted. It also states which parts of the market were never sampled.
A buyer asking "Which CRM works for a five-person field service team?" creates a different measurement from "Tell me about Brand A." The second question can check description accuracy. It cannot establish that an assistant would independently recommend Brand A to that buyer.
Four outcomes, four counting rules
| Measure | Numerator | Denominator |
|---|---|---|
| Mention rate | Valid answers naming the brand | All valid answers in the defined panel |
| Recommendation rate | Valid answers positively suggesting the brand for the stated need | All valid answers in the defined panel |
| Owned-site citation rate | Valid answers linking to the brand's domain as a source | All valid answers in the defined panel |
| Share of voice | Counted appearances of the brand | Counted appearances of all brands in the declared comparison set |
For share of voice, count each brand at most once in each answer. Publish the comparison set and say whether it includes every brand named or a fixed list of rivals. Those are different measures and should not share an unlabeled trend line.
Illustrative example: 100 valid answers name Brand A in 20, positively recommend it in 10, and link to its domain in 5. Mention rate is 20%, recommendation rate is 10%, and owned-site citation rate is 5%. If the declared brand set has 80 total counted appearances, Brand A's share of voice is 20/80, or 25%. None of those numbers is market share or attributed revenue.
An answer warning a buyer away from Brand A counts as a mention and not a positive recommendation. A link to Brand A's comparison article can coexist with a recommendation of Brand B. Preserve those distinctions during human review and automated classification.
Build the collection contract before the dashboard
Record these fields for each measurement cycle:
- Exact prompt wording and panel version, with the buyer need each question represents.
- Engine, product surface, model/version where exposed, search mode, and collection date.
- Country or market, language, session conditions, and relevant personalization settings.
- Planned repetitions and actual valid answers per prompt and engine.
- Failed requests, refusals, empty answers, and other exclusions, with counts and reasons.
- Saved answer text, linked sources, competitor names, and the classification decisions.
A transport failure is missing data, not an answer that omitted the brand. A refusal is an observed response; either include it under a documented rule or report it in a separate category. Do not silently retry failures until a preferred answer appears.
Consumer interfaces and APIs can use different retrieval, context, and product behavior. Label them separately. A clean session reduces some context effects, but does not make collection identical across users or products.
Repeat enough to answer the question you are asking
Small panels can reveal inaccurate descriptions or obvious omissions. Estimating an appearance probability with a narrow uncertainty interval requires more repeated observations and an explicit statistical model. The sampling guide explains the assumptions behind a roughly 100-run estimate for one fixed prompt and engine.
A probability that an entire brand list repeats is different from the probability that one brand appears. Do not use one statistic as a substitute for the other.
Compare periods without hiding changes
Keep the panel and aggregation rules stable. If prompts are weighted, publish the weights and what they represent. Equal prompt weights express the analyst's chosen panel; they do not establish real buyer query volume.
Report results by engine and prompt group before combining them. Flag model updates, changed search modes, new markets, revised classifications, and altered competitor sets. A substantially changed collection contract starts a new series or needs a documented overlap comparison.
A higher mention rate after a deployment is an observation. To claim that the deployment caused it, account for product updates, competitor changes, seasonality, and other explanations with an appropriate design.
Sources, traffic, and implementation
Google's AI search guidance describes indexing and retrieval foundations. OpenAI's search documentation explains current-source answers, source links, and access conditions. These establish product behavior; they do not supply a universal brand-ranking formula.
Track AI referral traffic and qualified inquiries alongside the answer panel. They answer different questions. Source links can suggest where factual or content corrections belong, while an inferred cause remains a hypothesis until tested.
Etched uses this distinction to connect measurement with implementation. A paid audit should agree the collection contract before running it. The free checker asks for a brand name and email and provides a narrower starting point.
Frequently Asked Questions
Is a brand mention the same as a source citation?
No. A mention names a brand in an answer. An owned-site citation links to its domain as a source. A positive recommendation suggests the brand for the buyer’s need. Count the three separately.
What should the denominator include?
State the valid answers used for each rate, along with failed, refused, and excluded responses. Share of voice uses counted brand appearances in a declared comparison set, not the answer count.
Can I compare API collection with a consumer interface?
Treat them as separate collection surfaces unless their equivalence has been established. Record model, search mode, session context, market, and date for both.
Does a visibility increase prove the work caused it?
No. A before-and-after change can have several explanations. Retain the change log and use a suitable comparison or experiment before making a causal claim.
Last updated: October 15, 2026
