How to Measure AI Search Visibility, Citations and Leads
Use a fixed question sample, observed source links, measurable referrals and qualified enquiries without combining unlike outcomes.
Measure AI search visibility at several stages: brand appearance, linked citation, measurable visit and qualified enquiry.
- Use fixed branded and unbranded questions in fresh sessions.
- Report counts and denominators for each platform, language and round.
- Keep Google reporting, cross-platform observations and analytics attribution separate.
Measure AI search visibility at several stages: whether the brand appears, whether an answer links to a site page, whether a visitor reaches the site, and whether the visit becomes a useful enquiry. Keep these outcomes separate. A repeated question sample can show changes in observed answers, while platform reports and analytics describe different parts of the journey.
Define mention, recommendation, citation, referral and enquiry
| Outcome | Definition | What it does not prove |
|---|---|---|
| Mention | The brand or person is named | A site URL was used as a source |
| Recommendation | The answer explicitly presents the service as an option | The recommendation was clicked |
| Linked citation | A working site URL supports the answer | A visit or enquiry occurred |
| Referral | Analytics records a visit from an identifiable source | Every platform visit is attributable |
| Qualified enquiry | A genuine request fits the offered work and project scope | Every successful form event is qualified |
Choose questions from real audience decisions
Build the sample from questions prospective clients ask when diagnosing a problem or choosing help. Separate branded controls from unbranded informational and commercial questions. Assign each question an editorial owner—such as a guide or service page—for content review, but never append that URL to the prompt sent to the assistant.
Keep the core set stable across rounds. Optional diagnostic questions can explore a change, but report them separately so they do not silently alter the baseline.
Record platform and search context
For every run, record the platform and visible model/product label, whether search was available and actually used, timestamp, language and region context, run status, mention/recommendation/citation coding, exact source URL and a short evidence reference. Use fresh sessions because answers can vary with context and product changes.
An answer produced without search is still an observed behavior. Record it rather than deleting it. An unavailable run is unknown, not a zero-citation answer.
Calculate rates with one declared denominator
For a citation-bearing-run rate, divide successfully observed unbranded runs containing at least one qualifying on-site source link by all successfully observed unbranded runs in the same platform, language and round. Multiple site links in one answer still count as one citation-bearing run for this rate.
Illustrative example—not this site’s results: 8 of 60 successfully observed unbranded runs include a qualifying site source. The observed citation frequency is 8 ÷ 60 = 13.3%. If six other runs failed, report those six separately. The result is not a share of every question asked on the platform.
Report mentions and recommendations separately and show both count and denominator. The AI citation readiness guide explains how to investigate access and evidence before interpreting a finite sample.
Use Google’s Generative AI performance report within scope
Google documents a dedicated Generative AI performance report in Search Console for eligible AI Overviews and AI Mode visibility, with dimensions such as page, country, device and date. Google also explains that this activity is represented in Web performance data. Do not add the dedicated report’s impressions to Web totals as incremental traffic, and do not treat it as ChatGPT, Claude or Perplexity reporting. See Google’s Generative AI performance report definition.
If access or indexing needs diagnosis before interpreting the report, use a Google Search Console indexing review. A documented report’s existence does not prove that a particular property has data.
Trace only measurable referrals
Use the existing analytics setup to review session source/medium and landing pages for observed referral hosts. Consent choices, stripped referrers, in-app browsers and later return visits can make attribution incomplete. Do not assign an AI source to direct or unknown traffic without evidence, and do not add invasive tracking to recover what cannot be observed responsibly.
Separate completed forms from qualified opportunities
A successful form response can emit the existing consent-aware lead_form_success event. That event indicates a completed submission, not automatically a relevant opportunity. Assess qualification privately using the requested service, genuine scope and follow-up outcome. Never send names, email addresses, messages or raw query text into analytics.
Use a monthly review to make decisions
- Freeze the platform/language question sample and record unavailable runs.
- Compare mention, recommendation and linked-citation counts using the same denominators.
- Review cited landing pages and whether they answer the intended decision.
- Compare known referrals and completed lead events without inventing attribution.
- Review qualified enquiries privately.
- Choose one evidence, access or page-ownership improvement for the next cycle.
A rise in citations may justify strengthening the cited page and its service path. Stable access with no citations may point to relevance or evidence gaps. More referrals without qualified enquiries may indicate a landing-page or offer mismatch. None of these observations alone proves causation.
Connect measurement with the pages that matter
The AI Search Optimization service can connect question ownership, source-page readiness and a measurement method with known limits. Learn more about Murat’s Technical SEO work, or share the website, target market and questions prospective clients ask.