AI Search Visibility:
Measure Mentions and Citations

AI search visibility is a brand’s presence in AI-generated search answers, measured here through brand mentions and citations. Mention rate measures how often the answer names the brand; owned citation rate measures how often it links to a brand-controlled website as a source. Calculate both against a defined set of observed answers.

The useful starting point is a record of the answers themselves: which question was asked, which search experience answered it, what name appeared, and which pages were cited. A percentage becomes interpretable only when its counting rules and denominator are available alongside it.

What AI search visibility measures#

A brand mention is an identifiable appearance of the brand in the answer text. An owned citation is a source link to a page on a domain included in a predefined ownership list. A third-party page discussing the brand belongs in a separate citation category; it does not become an owned citation because it names the brand.

These signals can be counted independently. In Ahrefs’ Brand Radar definitions, repeated appearances of a brand within one response count as one mention, and multiple cited pages from one domain count as one domain citation. Those product definitions provide a useful reference for response-level counting. The method below specifies its own rules so that a manually collected report can be reproduced.

Keep the search experience explicit. According to Google’s AI features documentation, AI Overviews and AI Mode may use different models and techniques, producing different responses and supporting links. Treat those experiences as separate measurements even though both belong to Google Search.

Visibility also differs from several adjacent observations:

• Retrieval: a page was found or considered during answer generation.
• Recommendation: the answer explicitly suggests choosing a brand or option.
• Position: a name or citation occupies a recorded place within an answer.
• Traffic: a visit reaches a website.
• Conversion: a recorded action occurs after a visit.

For retrieval specifically, Ahrefs separates cited pages from pages found without being cited. A crawl or retrieval record is therefore insufficient evidence for the visible citation count defined here. Similarly, a mention count alone does not distinguish a recommendation from a criticism; preserve the surrounding text for that assessment.

How to calculate mention rate and owned citation rate#

Use one completed, inspectable answer as the counting unit. For each answer, record two binary flags: whether the brand is mentioned in answer text and whether an owned page is visibly cited. Count each flag at most once per answer, regardless of repeated names or multiple owned links.

Let N be the number of eligible observed answers, M the number containing a brand mention, and C the number containing at least one owned citation.

Mention rate = M ÷ N × 100%

Owned citation rate = C ÷ N × 100%

Show M, C, and N alongside the percentages. When N is zero, report the rates as unavailable. These formulas describe the collected answers under the stated conditions; they do not estimate the percentage of all AI users who saw the brand.

Define the denominator before collecting results#

For this method, an eligible answer is a completed response from the designated AI search experience with enough saved evidence to inspect its answer text and citations. Include completed answers without mentions or citations. Record collection errors, refusals, and unavailable answers separately, using the same exclusion policy in every reporting period.

Google states that AI Overviews do not trigger for every query. When monitoring that feature, record queries without an overview separately from collection failures. Report both the number of completed searches and the number that produced an inspectable overview. A citation rate calculated only across generated overviews has a narrower denominator than a rate calculated across all completed searches; label the denominator explicitly.

Do not restrict N to answers that already contain a citation. That would measure presence within citation-bearing answers and omit completed answers with no sources. It would be a different metric from the response-level rate defined above.

Record the overlap#

Let B be the number of answers containing both a brand mention and an owned citation. The four outcomes partition the same set of eligible answers:

These counts should add up to N. Report the share of answers with either signal as (M + C − B) ÷ N × 100% only when that combined measure is useful. Adding mention rate to citation rate without subtracting the overlap double-counts answers containing both.

Keep accuracy separate from presence. Under this method, an identifiable but inaccurate brand description still counts as a mention, with an additional accuracy flag. This preserves the visibility count while making incorrect descriptions available for review.

Build a repeatable measurement process#

Freeze a prompt panel for each comparison period. Store exact wording and group prompts by topic and purpose: information seeking, problem solving, comparisons, and questions naming the brand. Keep branded prompts separate from prompts that do not supply the brand name, because they measure presence under different conditions.

Choose prompts from documented search and content questions, and record why each belongs in the panel. Describe the panel as a monitored sample. Adding or removing prompts changes what its aggregate rate measures; establish a new baseline when the panel changes.

Repeat prompts rather than relying on a single answer. A preprint on generative search measurement reports citation variability across repeated queries on three platforms and three consumer-product topics. Its findings support treating observed visibility as an estimate with uncertainty, while its limited sample does not establish a universal number of runs.

Use the same planned repeat count for each prompt within a reporting segment. Keep engine, mode, language, location, account settings, and conversation context as consistent as the collection method permits. Record conditions that cannot be controlled. Start fresh conversations for a panel intended to measure initial answers; measure follow-up conversations separately.

For each run, save:

• The exact prompt, prompt identifier, and topic group.
• The platform, search feature or mode, and model label when available.
• The timestamp, language, location settings, and relevant account or session conditions.
• The complete answer text and its displayed source links.
• The mention flag, owned citation flag, matching evidence, and accuracy assessment.
• The run status, including errors, refusals, and absence of the monitored AI feature.

Specify brand matching before scoring. Maintain approved names and aliases, distinguish the brand from unrelated uses of the same word, and decide whether product names count toward the parent brand. Under the definition here, a name appearing only in a source title does not count as a mention in answer text.

Specify citation matching just as carefully. Maintain the owned-domain list, decide how subdomains and redirects are handled, and save the cited URL. Count links explicitly presented as supporting sources. Keep additional retrieved pages outside that citation flag, and apply the same treatment of inline links and source panels across periods.

Calculate rates first by platform and topic. A combined rate weights segments according to their share of collected answers; changing that mix can change the total even when each segment is unchanged. Retain segment results and equal repeat counts when comparing periods. Repeated runs of the same prompt are repeated observations of that question, not additional unique questions.

Which tools can measure AI search visibility?#

Manual answer collection provides direct evidence for the two response-level rates. A monitoring tool can automate that collection, but its exported data and counting rules determine whether its dashboard matches this method. Before comparing results, check the platforms covered, prompt corpus, collection conditions, repeat schedule, brand matching, domain ownership rules, and denominator.

First-party webmaster reports provide another view of visibility. Their reporting units differ from a manually monitored answer panel.

Google Search Console#

Google’s generative AI performance report announcement describes dedicated views of URL impressions within AI features in Search and Discover, with page, country, date, and Search device information. The announcement notes worldwide rollout to all websites on August 31, 2026.

Use these reports to examine which pages receive Google AI feature exposure. Their unit is an appearance of a URL within Google’s reporting system, rather than a brand mention in a saved response. An impression total does not supply N for the response-level formulas above, and the report does not cover appearances on other AI platforms.

Bing Webmaster Tools#

Microsoft’s AI Performance launch documentation describes citation reporting across Microsoft Copilot, Bing AI-generated summaries, and selected partner integrations. It includes citation totals, cited-page information, and sampled grounding queries: phrases used to retrieve content associated with citations. Microsoft states that citation counts do not establish ranking, placement, or page importance.

Use page-level citation activity to identify URLs for closer inspection. Grounding queries describe retrieval activity; treat them separately from the original user prompts stored in a monitoring panel. The supported Microsoft and partner coverage is also narrower than all AI search activity.

Bing’s Citation Share documentation defines a separate metric: citations attributed to a site divided by citations across all sites for the same grounding query. Microsoft describes it as an observational measure rather than ranking, traffic share, or a content quality score. Its denominator is citations, while the owned citation rate above uses answers.

Keep Google impressions, Bing citation totals, Bing Citation Share, and monitored answer rates in separate columns. Compare each against an earlier period using its own definitions.

How to improve AI search visibility#

Begin with access and eligibility, then assess the content against the questions in the prompt panel.

For Google, supporting-link eligibility requires a page to be indexed and eligible to appear in Search with a snippet. Google recommends ordinary SEO practices, accessible text, internal links, and structured data consistent with visible content. It says special AI files or markup are unnecessary, and meeting its requirements does not guarantee inclusion.

For ChatGPT search, the official OpenAI crawler documentation identifies OAI-SearchBot as the search crawler and recommends allowing it through robots.txt and the published IP ranges. It distinguishes these search controls from GPTBot’s training controls: allowing search access and allowing training are independent settings.

After checking the relevant platform’s access requirements, review the saved answers and associated pages. Use the following checks to select an edit:

• Question coverage: identify which monitored questions the page answers and which important details remain absent.
• Source identity: make the publisher, organization, or product identity explicit and consistent with the matching rules.
• Evidence: support factual claims with accessible sources and dates where relevant.
• Accuracy: correct outdated information on controlled pages and record incorrect AI descriptions separately.
• Structure: organize the answer under descriptive headings, with tables or lists where they clarify the subject.

These are content and access checks, not a forecast of citation gains. Record the URL, the change, and its date, then repeat the unchanged prompt panel. Assess which answers changed and whether that pattern persists across runs before interpreting an aggregate increase.

What an AI visibility report should contain#

Report the prompt-panel version, collection period, platforms, settings, repeat policy, completed runs, and exclusions. Show mention and owned citation counts with their denominators, the four overlap outcomes, and results by platform and topic. Include the saved answer evidence behind notable changes.

Use the same panel and matching rules for comparisons between brands. Compare performance against a baseline collected under those conditions; a percentage from a different prompt corpus or product definition cannot establish a pass mark for this panel.

Keep referral visits and recorded conversions in separate analytics measures. The response-level rates establish appearances in collected answers; they contain no click or transaction event. Evaluating a business outcome requires evidence for that outcome beyond the visibility count.

The report’s central result should remain specific: how often the brand was named, how often an owned page was cited, where those events overlapped, and which questions produced them. That gives AI search visibility a clear meaning and provides the evidence needed to choose the next page or access issue to address.