INSIGHTS / SOURCES & MEASUREMENT

How to Measure AI Visibility: Mentions, Citations, and Outcomes

Measure AI visibility against a declared sample and a clear counting rule. Then report identifiable visits, qualified inquiries and revenue separately. A tested answer is not a consumer, a search impression or a sale.

8 min read

TL;DR

  • Define the prompt sample, platforms, conditions and eligible-answer rules before collecting results. Keep failures and exclusions visible.
  • Calculate mention, owned-domain citation and recommendation rates separately. They can overlap, and repeated answers are not independent consumers.
  • Check source support and factual accuracy alongside visibility. Keep identifiable referrals, qualified inquiries and revenue separate.
  • Compare equivalent samples and disclose weighting. Platform impression reports and your own answer sample measure different things.

Define the unit and the sampling plan.

AI visibility describes observations of whether and how a brand or its sources appear in artificial intelligence (AI) answers. There is no single universal visibility score: define the product, question set and event you are measuring.

For the method below, one unit is one complete, eligible answer from a planned prompt run. These are editorial reporting definitions, not universal platform metrics. Freeze the prompts, clusters and status rules before collection.

Keep branded and unbranded questions separate. Record platform, product mode, language, market and session conditions. If a setting is unknown, record unknown; do not infer it. Keep repeat runs so you can see variability, but do not treat repeats of the same prompt as independent people.

Predefine how to classify failures. Our example protocol counts a complete substantive answer, including one with no brand mention or citation. It records a complete answer saying no relevant results were found as completed, with zero mention/citation. A network failure or timeout is a failed run. A pure refusal is a separate refusal status and excluded from the example’s eligible-answer rates. Publish each status count; choose and document your rules before seeing results.

Keep planned runs, completed responses, eligible answers, refusals and failures visible. Do not silently replace unfavorable answers with retries. If retries are permitted, preserve attempts and label which attempt contributes to the denominator.

Metric definitions and counting rules.

Here, citation rate means owned-domain citation rate per eligible answer. Count an answer at most once for each rate, even if it repeats the brand or links several owned pages. Brand aliases and owned domains must be agreed before labeling. Recommendation means the answer presents the brand as a suitable option for the stated need; a neutral mention or a rejection does not qualify.

Metric definitions and counting rules.
MetricUnit and calculationDeduplication / sourceCannot establish
Brand mention rateEligible answers mentioning the brand ÷ all eligible answersOnce per answer; saved text and agreed brand aliasesRecommendation, impressions or market share
Owned-domain citation rateEligible answers linking an owned domain ÷ all eligible answersOnce per answer; resolve owned domains and retain URLsA click or factual support for every claim
Third-party citations mentioning the brandEligible answers with a cited external page whose text mentions the brand ÷ all eligible answersOnce per answer; inspect external text, record unavailable sources as unknownAn owned-domain citation or an endorsement
Recommendation rateEligible answers meeting the pre-agreed suitability rule ÷ all eligible answersOnce per answer; human label and retained contextA purchase decision
Distinct cited owned URLsCount unique normalized owned page URLs in the sampleDocument normalization; remove tracking parameters, preserve meaningful variants and raw linksNumber of answers, visitors or leads
Verified factual errorsCount distinct claim–verified-fact conflicts; also report affected answersGroup repeated instances of the same conflict; retain all occurrences and approved evidenceA complete inventory of every possible error
Identifiable AI referral sessionsAnalytics sessions classified by observed AI referrer under a saved ruleUse the analytics session definition; filter tests and bots under declared rulesAll AI influence; unattributed visits remain unknown
Qualified inquiriesAccepted inquiries meeting agreed market, need and service-fit criteriaGroup duplicates into an opportunity; retain submissions, test/spam labels and qualification reasonsBooked revenue or proven GEO causality

Worked calculation: 200 answers, three overlapping rates.

Simulated data, not a client result: 50 fixed prompts × 2 platforms × 2 repeats = 200 planned runs. Assume all produce complete eligible answers, with no failures or exclusions. Forty-eight mention the brand, 30 cite an owned domain and 20 meet the recommendation rule.

Worked calculation: 200 answers, three overlapping rates.
MeasureCalculationResult
Mention rate48 ÷ 20024%
Owned-domain citation rate30 ÷ 20015%
Recommendation rate20 ÷ 20010%

The three sets can overlap, so do not add the percentages. An answer can mention, cite and recommend the same brand. These rates describe this sample, not exposure among real users. The 200 runs are not 200 independent consumers.

For a change from 20% to 30%, the absolute difference is 10 percentage points; the relative increase is (30 − 20) ÷ 20 = 50%. With a zero baseline, relative growth is undefined: show counts and percentage-point change instead.

A sampling record you can copy.

Keep one row per planned run and an immutable reference to the saved answer. Use protected storage for full transcripts; do not include private customer details in a public report. The template is a reporting aid, not an automated platform integration.

A sampling record you can copy.
Field groupHow to fill it
run_id / prompt_id / clusterUnique run and stable prompt IDs; preserve exact prompt text in a separate prompt register
branded_flagBranded or unbranded, assigned before collection
platform / mode / language / marketActual product and known settings; use unknown for unavailable fields
timestamp / session_conditionsTime with timezone; new or continuing session; relevant personalization state if known
run_status / answer_referenceCompleted, failed, timed out or refused; link to preserved evidence
brand_mentioned / owned_domain_citedYes/no for eligible answers; not assessed for failed runs
citation_urls / recommendation_labelAll raw links; yes/no/needs-review under the agreed recommendation rule
accuracy_notesClaim, approved fact, evidence link, reviewer decision; unresolved items remain unverified

Sampling record fields

run_id,prompt_id,cluster,branded_flag,platform,mode,language,market,timestamp,session_conditions,run_status,answer_reference,brand_mentioned,owned_domain_cited,citation_urls,recommendation_label,accuracy_notes

Review accuracy as well as presence.

A citation link is not proof that the source supports the answer. Check the linked passage, date, subject and scope; label supported, unsupported or unverified. The citation diagnostic table shows common failure modes.

When two reviewers disagree about a recommendation or error, retain the disputed label and resolve it against the written rule. Do not quietly change the rule to improve a rate. Report how many labels remain unresolved and whether they affect a numerator.

Separate visibility from attribution.

A higher citation rate does not prove that a later inquiry came from that citation. An observed referrer is one clue. A buyer’s self-reported source adds context but is not perfect causal evidence. Missing referrers do not establish absence of AI influence, and you cannot assume assistants preserve a tracking parameter you append.

A call-to-action click is not a submission. Count confirmed saved requests, exclude labeled tests and spam from qualified totals, and preserve duplicate records while grouping related requests under your stated rule. Record revenue only from actual business records, with a separate attribution method and uncertainty.

Keep display data, sampled answers, identifiable sessions, submitted requests, qualified opportunities and revenue in separate report rows. They may be related, but they do not form an observed end-to-end funnel without matching evidence.

Compare equivalent samples and state the weighting.

Hold prompts, markets, language, mode, session rules and platform mix constant. Show additions separately. If a product or protocol changes, annotate the break rather than presenting a clean before-and-after comparison.

With unequal platform samples, report each platform first. A pooled rate weights platforms by their eligible-answer counts; an equal-platform average weights their rates equally. Equal-cluster averaging is different again. State which you use.

Illustrative example: platform A has 10 mentions in 100 answers (10%), while B has 10 in 20 (50%). The pooled rate is 20/120 = 16.7%; the equal-platform mean is 30%. Both calculations are valid, but they answer different questions. Neither is a market-wide estimate.

Review failed and excluded runs beside these rates. A smaller denominator after outages may change the result even if the underlying brand visibility did not.

What current platform reports can tell you.

As checked on October 5, 2026, Google documents a Generative AI performance report (Search) for AI Overviews and AI Mode impressions. Its dimensions include pages, countries, dates and devices; Search Labs experiments are excluded. Current report documentation.

Chart totals aggregate by property unless a URL filter changes that basis. Page rows aggregate by page, so totals can differ. The report overlaps Web performance data: do not add both as independent exposure. The documentation describes impressions, not complete prompts, brand-mention rates, other AI platforms or conversion attribution.

Google notes worldwide rollout but retains access/low-impression caveats. Verify what your property shows; absence is not a clean zero. Recent data can be preliminary, and row/export limits apply. Its more general AI features page still describes inclusion in Web data; that alone is no longer the full reporting picture.

End the report with a decision.

Name the question cluster, saved observation, evidence gap, responsible role and next check. Continue a useful task, adjust a weak test, investigate a source conflict or stop work with no defensible business purpose. The strategy guide turns findings into implementation stages.

Bring your offer, markets, proposed prompts and current reporting definitions when discussing SemShift’s AI visibility audit. A reproducible starting point is more useful than an unexplained visibility score.