AI visibility metrics describe what happens in a defined set of AI answers. To make them useful, count brand mentions, links to your website, recommendations, and competitor appearances separately. Report the sample behind each number, then connect the findings to the pages and buying decisions you want to improve.
A dashboard percentage without a denominator is difficult to interpret. A brand appearing in half of ten branded questions means something different from appearing in half of two hundred unbranded discovery responses across two countries.
This guide explains a practical measurement system for marketing teams serving the United States and Canada. It includes a worked fictional dataset, formulas you can reuse, a reporting template, and a way to move from an observation to a content decision.
Define the response set before choosing a metric
A prompt is the question you track. A response is one completed answer to that question under a particular set of conditions. Repeated runs of the same prompt produce additional responses, not additional unique buyer needs.
Record the exact prompt, country setting, language, platform, available model label, collection time, and whether search was used or visible. Give each response a unique identifier so you can review the original answer later.
Keep failures separate. A timeout or an unsuccessful collection is not an answer where the brand was absent. Show the number of attempted runs, completed responses, and failures before calculating visibility rates.
Define the business scope too. Your measurement might concern one product, a service area, or a buyer journey. A clear scope helps explain why the selected questions matter and which decisions the results can support.
Create rules for recognizing the brand
Write down the official business name, accepted variations, and domains that belong to it. Include a procedure for ambiguous names. A reference to another company with a similar name should not count as your visibility.
Decide how to handle product names, parent companies, and acquired brands. If a product can appear without the corporate name, keep those observations distinguishable rather than silently combining them.
For a local business, location can help identify the right entity. For a software company, a linked domain or a specific product description may resolve ambiguity.
Have a reviewer inspect uncertain cases. Consistent classification is more valuable than pretending every response fits a simple automated rule. Save the reason for corrections so future reviewers can apply the same interpretation.
Calculate brand mention rate
Brand mention rate answers a narrow question: in what share of completed responses did the intended brand appear at least once?
The formula is completed responses mentioning the brand divided by all completed responses in the defined set, multiplied by one hundred.
Count a response once even if the name appears five times. Otherwise, a verbose answer can contribute more to the rate than a short answer serving the same buyer need.
Keep positive, negative, and neutral context available for review. A mention establishes presence. It does not by itself establish endorsement, accuracy, or purchase intent.
For decision making, break the rate down by prompt family. Discovery questions reveal a different kind of presence from questions that already supply the business name. An overall rate should not conceal that difference.
Calculate owned citation rate
Owned citation rate measures how often a completed response includes at least one visible citation to an approved domain you control.
The formula is completed responses citing an owned domain divided by all completed responses in the defined set, multiplied by one hundred.
Record the cited URL as well as the domain. A link to an obsolete page, an irrelevant resource, or a different regional version may require a different action from a link to the intended guide.
Do not confuse a visible citation with every source that may have influenced an answer. Your metric describes the links you can observe. Label it accordingly.
Keep total citation counts in another field if needed. One response can contain several links to your website while contributing only one successful response to the owned citation rate.
Separate outside coverage from owned citations
An independent review can mention your company even when the answer does not link directly to your website. That can be relevant coverage, but it belongs in a different field from an owned citation.
Record the outside publisher, page URL, brand reference, and role of the source. It may support a comparison, establish a location, or describe a particular capability.
This distinction improves the action list. An owned content gap may be something your editor can address immediately. An inaccurate outside profile may require a correction request or an update to a listing you control.
Avoid treating the appearance of an outside source as a permanent preference for that publisher. Review its contribution to the specific question before investing in a new coverage campaign.
Define recommendation rate carefully
A recommendation needs a classification rule. For example, count a response when it presents your product among options suitable for the buyer's stated need.
A passing reference in a warning, a historical example, or an explanation of an unrelated category would not meet that rule.
Save the surrounding sentence and any important qualification. A product recommended only for a different audience should not be reported as an unconditional fit for your target buyer.
The formula can mirror mention rate: completed responses meeting your recommendation rule divided by completed responses in the relevant set.
Use this metric only where recommendation is a sensible outcome. An educational prompt about calculating a percentage may not call for vendors at all. Mixing those responses into a recommendation target can obscure the purpose of the test.
Choose a share of voice definition
Share of voice has no useful meaning until you state what is being shared. One practical definition uses brand appearances across a fixed comparison set.
Count each selected brand at most once per completed response. Add those appearances across the response set. Divide your brand's appearances by the total appearances of every brand in that fixed set.
This produces a share of selected brand appearances. It is not the percentage of all responses, the share of all possible competitors, or a measure of market revenue.
A response naming three selected brands contributes three appearances to the denominator. A response naming none contributes zero appearances, although it remains part of the mention-rate denominator.
Keep a separate no-brand response count. Without it, share of voice can look strong while most answers do not discuss any of the companies you monitor.
Work through a fictional twenty response example
Imagine a fictional business called Northline and two selected competitors, Cedar and Harbor. The following counts are illustrative, not observations from any customer account.
The team completed twenty responses. Northline appeared in eight, Cedar in ten, and Harbor in six. Because several answers named more than one business, the twenty responses contained twenty-four selected brand appearances.
Northline received five recommendations under the team's stated rule. Its owned domain was cited in three responses. Four responses contained no selected brand. Two of the three responses citing Northline's domain did not name Northline in the answer text.
These separate fields let the team distinguish a cited educational resource from a product recommendation. Review the response records before calculating percentages so overlapping events and ambiguous names are classified consistently.
The reviewed calculations are:
- Northline mention rate — Calculation: 8 divided by 20 — Result: 40 percent
- Northline owned citation rate — Calculation: 3 divided by 20 — Result: 15 percent
- Northline recommendation rate — Calculation: 5 divided by 20 — Result: 25 percent
- Northline share of selected appearances — Calculation: 8 divided by 24 — Result: 33.3 percent
- Responses with no selected brand — Calculation: 4 divided by 20 — Result: 20 percent
A complete dataset can have these totals even though the figures overlap. Recommendation, mention, and citation are separate observations about a response. Do not add their percentages as though they were mutually exclusive categories.
Review the example without hiding the denominator
The 40 percent mention rate and 33.3 percent share of selected appearances answer different questions. The first uses completed responses; the second uses appearances of three selected brands.
Changing the comparison set changes share of voice even when your own appearances stay constant. If a newly added competitor appears frequently, your share can fall without your brand disappearing from a single answer.
Changing the prompt mix can also change the rates. Adding more branded questions may increase presence while leaving discovery untouched.
For each reporting period, preserve the prompt version, competitor set, countries, platforms, and collection schedule. If any changes, annotate the break and retain a comparable subset where possible.
Report country and platform differences
Keep US and Canadian results visible. Different product availability, service areas, currency, and local requirements can make the underlying questions materially different.
When combining compatible samples, calculate the overall rate from the combined counts. Do not average percentages without accounting for their denominators.
For example, two mentions in four Canadian responses and eight in sixteen US responses combine to ten mentions in twenty responses, or 50 percent. If the sample sizes differ and the rates differ, the simple average can mislead.
Break platforms out as well. A combined number can be a convenient summary, but it should not hide that one platform improved while another declined.
Use market weighting only when you have a reason for it. Label the weights as a business choice rather than presenting them as a natural property of the response data.
Keep position and sentiment in context
Position can describe where a brand appears in a list or answer. Define the unit before comparing it: list order, first textual occurrence, or another documented measure.
A second position in a two-option answer is not directly comparable with second position in a ten-option answer. Record whether the response actually presents an ordered ranking.
For sentiment, separate the classifier's label from a reviewer's assessment of factual accuracy. A positive description can still contain an incorrect feature claim. A qualified description may be accurate and useful to the buyer.
Use these measures to find responses worth reading. A change in the label should lead to inspection of the underlying text rather than an immediate conclusion about brand reputation.
Connect each metric to an action
A missing owned citation can prompt a review of the page that should answer the question. A mention with incorrect capabilities can prompt a product fact correction. Repeated competitor appearances can prompt a comparison of use case coverage.
Not every observation requires new content. You may need to update an existing guide, clarify documentation, repair access, or correct a listing.
Give the action an owner, affected page, evidence, expected improvement, and review date. An expected improvement can be more accurate information or better coverage of a buyer question; it need not be an invented percentage target.
Keep the observation and action linked. When you return to the response set after an update, the team should understand what changed and why.
Use a compact reporting template
Start the report with the scope and sample, then show the metrics that support a decision. A useful template includes these fields:
- Business question and audience.
- Prompt set version and number of unique prompts.
- Countries, languages, platforms, and collection dates.
- Attempted runs, completed responses, and failures.
- Brand mention, owned citation, and recommendation counts.
- Fixed competitor set and share of voice definition.
- Important accuracy issues and source gaps.
- Page changes, responsible owners, and next review date.
Include a short evidence table with selected response identifiers and URLs. The report should let another person reproduce the classification, not simply admire a percentage.
For an executive update, lead with the decision. For example, explain that the team will update an incomplete integration guide because the same buyer requirement recurred in several reviewed answers.
Connect visibility to visits and signups
Visibility measures what appeared in the answers you monitored. Analytics measures activity recorded on your website. Your CRM or product analytics can help evaluate whether those visitors became qualified leads, trial users, or active customers.
Keep those systems' definitions intact. A citation is not a website session, and a signup is not necessarily a paying customer. A person may also discover a business through one source and return through another.
Compare trends and investigate the relationship without multiplying a visibility percentage by an assumed audience to manufacture traffic estimates.
Make the landing page useful for the next decision. A reader learning the measurement method may need a platform explanation before onboarding. A reader already comparing tools may be ready to start a trial.
Organize the measurement workflow in SERPY
The SERPY platform connects the questions you monitor with the content work they reveal. AI Visibility provides tracked answers, sources, and competitor context for review.
Use Opportunity Discovery when the evidence points to missing coverage. Move a selected improvement into Content Engine, and use Performance Tracking to follow page and visibility signals.
Keep the metric definitions attached to your report, particularly when comparing results with another vendor's dashboard. A similarly named score can use a different denominator.
Try SERPY for free to begin with a focused set of buyer questions, inspect the answers behind the numbers, and choose a content improvement your team can carry through.
Frequently asked questions
What are AI visibility metrics
They are measurements of brand or content appearances in a defined set of AI responses. Useful examples include mention rate, owned citation rate, recommendation rate, and a clearly defined share of voice. Each should state its sample, classification rule, and denominator.
What is the difference between a prompt and a response
A prompt is the question being monitored. A response is one answer collected under recorded conditions. Ten repeated runs of one question provide ten response observations while still covering one unique prompt. Report both counts so repeated sampling is not mistaken for broader topic coverage.
How do I calculate mention rate
Divide completed responses that mention the intended brand by all completed responses in the defined set. Multiply by one hundred for a percentage. Count a response once, resolve ambiguous names, and keep failures outside the completed-response denominator.
What counts as an owned citation
Use a visible link to an approved domain you control as the classification rule. Save the exact URL and keep third-party brand coverage separate. Several links in one response can contribute to a citation count while contributing only once to the response-level citation rate.
Is a recommendation the same as a mention
No. A mention establishes that the name appears. A recommendation presents the offering as a suitable option under your stated rule. Read the context and preserve qualifications so an irrelevant example or unfavorable comparison is not counted as a recommendation.
What is AI share of voice
It is a share calculated within an explicitly defined comparison. One approach divides your brand's response-level appearances by all selected brands' appearances. Other methods can differ, so include the formula and competitor set instead of assuming every dashboard uses the same definition.
Can share of voice rise while mention rate falls
Yes. Your appearances can fall while competitors' appearances fall more sharply. The measures have different denominators. Review the underlying counts and responses before deciding whether the change represents a useful gain for your business.
Should I include answers with no brands
Include completed answers in mention-rate calculations even when no selected brand appears. Report the no-brand count separately for share-of-voice context. Excluding those answers from every metric can make a small amount of brand coverage look more extensive than it is.
Can I average US and Canadian percentages
Use combined counts for a combined rate when the samples are compatible. Preserve country breakdowns and explain material differences in prompts or requirements. An unweighted average gives each country the same influence regardless of how many responses it contributed.
How many observations establish a trend
There is no universal number that makes every selected prompt set representative. Use repeated observations under stable conditions and show the counts. Small samples warrant close inspection rather than confident claims about the entire category or a precise long-term change.
How do I connect the metrics to revenue
Review recorded website visits, qualified signups, activation, and customer outcomes alongside the visibility observations. Preserve the attribution rules in your analytics and CRM. A tracked citation can motivate investigation, but it is not itself evidence of a specific revenue amount.
How can SERPY help me use these metrics
Use SERPY to inspect the tracked answers, sources, and content opportunities behind a measurement. Keep your reporting definitions explicit and choose an action that serves a real buyer need. Try SERPY for free to connect that review with content planning and improvement.
