How to measure AI visibility
Measure AI visibility with defined mention rates, position, share of voice and returned-source share, plus a worked example and repeatable sampling plan.
How do you measure AI visibility?
Collect a fixed set of buyer questions under recorded conditions and save the complete answers. Define mention rate, position, share of voice and returned-source share before calculating them. Show each numerator and denominator, collection failures and reporting dates, then compare the same question groups over time.
Define what one observation means
To measure AI visibility, start with a record you can inspect: one answer to one question, collected through one interface at one time. Give that record an identifier. Save the full text, returned URLs, language, market, search mode and model when shown. Record collection failures alongside successful observations.
Choose questions from customer decisions and document why each belongs. Separate discovery, comparison and branded questions. Keep the wording fixed within a reporting period. If you change a question or collection method, mark the change and start a separate comparison group.
Use the definitions below as an explicit reporting convention. Publish them with your numbers so another reviewer can reproduce the calculation. Before adopting a tool's labels, use our AI visibility tool accuracy guide to check its saved answers and extraction rules.
Calculate mention rate from completed answers
Define a mention as an unambiguous reference to the target business in the answer text. Count the business once per answer, even if its name appears several times. Keep approved aliases and ambiguous names in a matching note. Review the surrounding sentence before crediting a match.
Mention rate equals answers naming the brand divided by completed eligible answers, multiplied by 100. Include completed answers that name no brands. Put refusals, empty responses, timeouts and unusable collections in separate status counts under your stated eligibility rule.
Show collection coverage beside the rate: completed eligible answers divided by scheduled observations. A missing response belongs in the collection record, while a completed answer without the brand belongs in the mention denominator. Preserve both counts when deciding whether the period needs additional collection.
Define position and its eligible sample
For this workflow, position means the order of distinct brands in an ordered recommendation list. Count a repeated brand once. Record “absent” when the brand is missing and “unordered” when the answer provides no meaningful recommendation order. Keep these statuses outside the average position calculation.
Report the distribution as well as the mean. Suppose a brand is first in three answers, second in five and third in four. Its position sum is 25, divided by 12 eligible answers, giving about 2.1. Show those twelve answers beside the total sample so readers can see how selective the average is.
Calculate AI share of voice within a named set
Define AI share of voice here as the target brand's mention events divided by all mention events for a fixed comparison set. A mention event is one brand appearing in one answer. Name the included rivals and keep that set unchanged during the comparison.
If the target appears in twelve answers, rival A in eighteen and rival B in ten, the set has forty events. The target's share is 12 divided by 40, or 30 percent. One answer can contribute several events because several distinct brands can appear together.
Keep newly discovered rivals in a separate note until the next planned set revision. If you expand the denominator, recalculate both periods using the same membership before comparing their shares. Keep mention rate alongside share of voice so readers can distinguish absolute presence in your sample from relative presence within the selected set.
Define returned-source share separately
For returned-source share, count each distinct normalized source URL once per answer. The same URL in two answers contributes two events. Preserve raw URLs, and document which fragments or tracking parameters you remove. Keep parameters that identify different page content.
Divide source events belonging to your domain by all returned-source events in the eligible answers. Match the actual hostname and document any included subdomains. Show how many answers had an inspectable source field and how many contained at least one URL; mark an unavailable field separately from an empty one.
A brand mention, returned URL and explicit citation are separate observations. Anthropic's web-search API documentation, for example, describes citation objects with source URLs and cited text. Preserve that evidence when available instead of inferring a citation from a URL alone. Anthropic's web-search tool documentation.
Work through a hypothetical example.com report
Imagine example.com represents a fictional booking business. You schedule eight questions four times each, using one collection method during a fixed week. Thirty answers complete and two attempts fail. All thirty have inspectable source fields; twenty-four contain URLs. The following figures are invented to demonstrate the calculations.
| Measure | Numerator | Denominator | Result |
|---|---|---|---|
| Collection coverage | 30 completed eligible answers | 32 scheduled observations | 93.75% |
| Mention rate | 12 answers naming example.com | 30 completed eligible answers | 40% |
| Average position | Position sum of 25 | 12 answers naming it in ordered lists | 2.1 |
| AI share of voice | 12 target mention events | 40 events across target, rival A and rival B | 30% |
| Returned-source share | 9 example.com URL events | 60 URL events across all source domains | 15% |
| Answers with returned URLs | 24 answers containing URLs | 30 answers with inspectable source fields | 80% |
The forty brand events come from the twelve, eighteen and ten counts above. The sixty source events use distinct URLs within each answer, so nine events for example.com could include one page returned repeatedly. Keep the underlying rows available for reconciliation, including the two failed attempts.
Choose sample size and cadence deliberately
As a manageable first plan, choose twenty distinct questions and collect each three times across a week. That schedules sixty observations per method and market. Review collection status weekly and summarize results monthly. Increase coverage where a decision needs more distinct questions or repeated observations.
Repeat collection because output can vary. Anthropic documents that identical API inputs can produce different results even at temperature zero. Anthropic's glossary on output variability. Keep scheduled repeats even when an early answer looks favorable, and record retries without silently replacing inconvenient results.
Show totals by question group and collection method before pooling them. In the example, failures leave some questions with fewer completed observations. A pooled rate weights completed answers equally; an average of question-level rates weights questions equally. Choose one, label it and preserve the calculation across periods.
Before the monthly review, open a selection of positive matches, absences and ambiguous names. Recalculate their rows from the saved answers. If you correct a matching rule, apply it to both comparison periods and record the revision. Keep the original labels available so another reviewer can understand why a previously published total changed.
For a zero denominator, display “not calculable” and the reason. That includes an empty comparison set with no mention events or a source sample containing no returned URLs. Preserve a genuine zero numerator when the denominator is positive; these two cases convey different information.
Explain changes and uncertainty together
These suggested sample sizes are work plans, not representative estimates of all users. Repeated answers to related questions are not independent market observations. Small samples, uneven failures and method changes can make movement inconclusive; six additional mentions in sixty completed answers changes the rate by ten percentage points. Before-and-after differences do not prove cause, and these metrics do not establish traffic or revenue.
SearchSeal monitors AI Overview, ChatGPT, Gemini and Perplexity today, saving tracked answers with brand mentions, position, rivals and returned source links. Explicit citation status is unknown. Its AI visibility workflow lets you inspect the saved evidence behind a result and use those observations in a report with clearly stated questions, dates and denominators.
Frequently asked questions
What should be in the measurement spreadsheet?
Include the question and answer identifiers, exact wording, date, method, collection status, matched brands, position eligibility and source URLs. Add the calculation rules and comparison-set membership in a separate methodology note.
How do I handle a new competitor?
Record it immediately as an observation. Add it to the formal share-of-voice set at a documented revision, then recalculate both comparison periods with the same membership if you want to compare them.
Should I count repeated mentions in one answer?
Under this guide’s convention, count each distinct brand once per answer. For source URLs, count each distinct normalized URL once per answer. Keep the two types of events in separate calculations.
How often should I review the numbers?
Use a weekly review to inspect collection failures and unusual answers, then a monthly review for the consistent question set. Record page edits and collection-method changes when they happen so they are available at the next review.