AI Search Visibility Explained: Mention rate, citation rate, and what each metric represents

AI search visibility is the observed presence of a brand or its content in a declared set of AI-generated answers. Mention rate measures how often eligible answers name the entity. Citation rate measures how often eligible answers visibly cite the declared domain or page. They are different events, and neither is traffic, ranking, authority, or causality.

For a fixed response set, use transparent response-level formulas:

mention rate = eligible responses with at least one qualifying entity mention
               / all eligible responses

owned citation rate = eligible responses visibly citing at least one declared owned URL
                      / all eligible responses

Here is an illustrative calculation, not company data. A fixed protocol produces 40 eligible responses. Twelve name the brand and six visibly cite its domain. Mention rate is 12/40; owned citation rate is 6/40. The six cited responses may overlap the twelve mention responses or occur elsewhere. The two rates cannot be substituted for each other.

There is no universal good mention or citation rate. Changing prompts, engines, locales, run timing, alias matching, or eligibility can move the fraction without any underlying improvement. Freeze those choices before comparing periods.

Mention rate measures answer-text presence#

Define the entity before collecting the numerator. List the canonical brand or product name, accepted aliases, ambiguous terms, spelling variants, and exclusions. Decide whether a parent brand and product count separately. Preserve the answer text so reviewers can see whether the match was real.

Ahrefs’ metric documentation counts a brand once per generated response even if it is repeated. That response-level rule answers a clear question: “In what share of eligible answers was the entity present?” Counting every textual occurrence answers a different question and can reward verbose repetition.

The mention rate alone says nothing about accuracy or stance. A recommendation, neutral comparison, caveat, and criticism all satisfy a basic presence rule. If those distinctions matter, code them separately with a reviewed rubric rather than changing the definition of mention after seeing results.

Citation rate measures visible source attribution#

A citation qualifies only when the answer interface visibly attributes an in-scope page or domain under the declared rule. Ahrefs separates a visible citation from a page merely “found in” the generation process. Background retrieval may be diagnostically useful, but it is not the event the reader sees.

The ownership boundary also needs a contract. Decide whether documentation subdomains, community pages, acquired domains, file hosts, and localized domains count as owned. Decide whether multiple owned pages in one response count once at response level or as several page-citation events. Use the response-level numerator for citation rate and report page counts separately.

Four cells reveal the relationship#

Mention and citation are two binary observations, so every eligible response occupies one cell:

Semrush’s study design separately coded brand mention and domain citation, demonstrating why the events cannot be collapsed. Its published rates describe its bounded sample and are not benchmarks for another prompt set.

The four cells support different questions. High mentions with low owned citation may show that third-party sources or prior model knowledge accompany entity presence. Owned citations without mentions may show informational source use. Neither pattern reveals why the model selected the content; it tells the team what to inspect in the raw answers and source set.

Define the denominator before the numerator#

An eligible response should satisfy the collection contract. Record:

• exact prompt and prompt version;
• prompt cluster and inclusion reason;
• engine, model or surface when observable;
• locale, country, language, and account state;
• timestamp and run identifier;
• whether web retrieval was available or requested;
• successful answer criteria; and
• exclusion rules for errors, refusals, empty answers, or unavailable sources.

Do not silently remove failures. Report attempted runs, eligible responses, and exclusion reasons. If one engine fails more often, a pooled denominator can conceal the problem. Calculate engine-level rates before any declared aggregation.

Semrush’s measurement guidance recommends a fixed prompt set for period comparison. That control prevents a larger or easier prompt roster from creating the appearance of improved visibility.

Repeated runs turn a snapshot into an estimate#

Generated answers can vary. A 2026 preprint used repeated sampling across its tested platforms and topics and found citation variability. Its sample does not prescribe a universal run count. It supports the narrower conclusion that one answer should not be reported as a stable population fact.

Repeat under a consistent protocol and show variation by prompt and engine. Preserve raw responses, not just the aggregate. When the run protocol changes, start a new comparable series or bridge the versions explicitly.

What the rates do not represent#

Bing’s AI Performance documentation is unusually clear: its citation counts do not indicate ranking, authority, performance, importance, placement, or role within an answer. Changes in citation volume are observational and cannot be attributed to one update from the report alone.

Keep these downstream or adjacent events separate:

Co-movement is useful for hypothesis generation. It is not enough for a causal claim.

A minimum reproducible report#

Keep the minimum metrics honest#

Mention rate answers whether the entity appeared in answer text. Owned citation rate answers whether the declared source appeared visibly. Their overlap shows how often both occurred in the same response. That small scorecard is useful because every numerator can be audited against raw evidence.