Loading AEO Improvement…

AEO Improvement field guide

How to Measure Visibility in AI Search Without Guessing

AI-search visibility is not one metric. A page can be indexed without appearing in an AI feature, cited without sending a visit, and visited without producing a qualified outcome. This guide gives you a repeatable, source-labeled worksheet to record each signal separately, compare like with like, and state only what the evidence supports.

· By AEO Improvement editorial

  1. 01 · Indexing
    Eligibility check
  2. 02 · AI reports
    Impressions or citations
  3. 03 · Visits
    Analytics sessions
  4. 04 · Outcomes
    Verified business events
Connect the records by page and date range. Keep their units and meanings separate.

Treat visibility as four records, not one score

A useful measurement starts by asking what each source actually observed. Search Console can report whether a page is indexed and, when the generative AI report is available, impressions in named Google Search features. Bing Webmaster Tools reports citation activity across its supported AI experiences. Analytics records visits that arrive at your site. Your CRM or commerce system records the outcomes your business defined.

Those records answer different questions. They are not interchangeable stages in a guaranteed funnel. A citation may not produce a click. A visit may not be attributable to an AI source. An outcome may follow several visits, a sales conversation, or work that happened outside the measured session.

  • Indexing: can this URL appear in a search feature that requires an indexed page?
  • Supported AI reporting: did a provider report an impression or citation for this URL or property?
  • Referral visits: did a session reach this page from a source that analytics can identify?
  • Business outcomes: did a defined lead, sale, or other goal occur within the attribution rules you chose?

1. Check indexing before interpreting visibility

Inspect the exact canonical URL in Google Search Console and record the verdict, inspection date, canonical information, and last crawl date when it is available. A missing crawl date is a missing observation, not a date to estimate. If you use a different search provider's index report, keep that result separate rather than treating one provider's status as universal.

Google says a page must be indexed and eligible to appear with a Search snippet to be shown as a supporting link in AI Overviews or AI Mode. Those checks establish technical eligibility only. They do not show that the page received an AI-feature impression or citation.

  • Record the URL Inspection result and the date you checked it.
  • Copy Google's reported last crawl date only when one is present.
  • Compare the selected canonical URL, not a redirect or a similar URL variant.
  • Keep the status beside the other evidence, but do not count it as an impression, citation, visit, or outcome.

2. Record supported AI impressions and citations separately

Google's June 3, 2026 Search Central announcement introduced separate Search and Discover generative AI performance reports. The Search report help page currently describes impressions for AI Overviews and AI Mode, with page, country, device, and date views. Discover has its own report. Google notes that the Search report may not appear when a property has not received enough generative AI impressions. Record the report and filters you actually used; do not turn a missing report into zero.

Google's AI-features guide also says AI Overviews and AI Mode activity is included in the overall Search Console Performance report under Web. Treat the dedicated view as a way to inspect supported AI-feature impressions, not a second batch of traffic to add to the overall total.

Bing Webmaster Tools uses a different unit. Its AI Performance report describes citation activity across Microsoft Copilot, AI-generated summaries in Bing, and selected partner integrations. It reports cited pages, citation counts, and grouped grounding queries. Bing's documentation explicitly says citations do not represent clicks, traffic, or engagement. The June 2026 additions for Intents, Topics, Citation Share, and Compare are described as preview capabilities, so mark them as previews if you use them.

Do not add Google's impressions to Bing's citations or compare their percentages as if their denominators matched. Keep the provider, feature, page, date range, geography, device, and available filters beside each number. Use the report's own definition and export rather than a manual sample of prompts.

  • Google: record the Search generative AI impressions report if it is available for the property. Keep Discover data separate.
  • Bing: record citations, cited pages, and grounding-query context from AI Performance. Do not label them impressions or visits.
  • If you run a prompt sample, label it as a sample. Record the prompt, date, location, engine, and setup; it is not provider-wide impression or citation volume.

3. Measure referral visits in analytics

In GA4, filter landing pages to the URL being measured and inspect sessions alongside the available session source, medium, and channel. Keep a consistent rule for which sources count as AI referrals. Google Analytics' current default AI Assistant channel includes sources such as ChatGPT, Gemini, DeepSeek, Copilot, and Grok, but explicitly excludes Google AI Overviews and AI Mode. A default channel name is therefore not a complete inventory of AI-origin visits.

Referrer information may be absent, grouped differently, or changed by a provider, browser, app, or redirect. Keep identifiable sources separate from direct or unattributed sessions. Do not assign every unexplained visit to AI search. If you use campaign tags on links you control, document the tagging rule and do not assume a tagged link identifies every organic AI citation.

  • Use sessions, not citation counts, as the visit measure.
  • Filter to the same destination URL and equal-length comparison periods.
  • Save the source and channel definition you used so a later report can be reproduced.
  • Treat unattributed traffic as unknown unless the evidence supports a source.

4. Connect business outcomes without overclaiming

Before collecting numbers, define the outcome that matters for this page: a qualified inquiry, booked consultation, trial activation, purchase, or another event your team can verify. Name the source system, event definition, attribution window, and whether the value is a count or an amount. If a lead becomes qualified in a CRM, use the CRM status rather than treating every form submission as revenue.

Compare outcomes for the same page and time windows where the data allows it. Record campaign changes, sales-cycle delays, tracking changes, and other known context. A higher citation count followed by a lead is a sequence, not proof that the citation caused the lead. Use attribution labels as the rules of a report, not as a causal experiment.

  • Write down the exact event or CRM stage before looking at the result.
  • Use the same attribution window and conversion definition in both periods.
  • Report counts and value with their denominator and scope. Small counts can move sharply in percentage terms.
  • Use language such as 'occurred during the period' unless the study design supports a causal claim.

Limits to keep on the record

Provider reports are bounded by the features, properties, dates, filters, and data each provider makes available. Google's report availability can depend on reported impression volume. Bing says its data is aggregated and is not a complete log of every time content may have been referenced; its grounding queries are generalized. A provider report is useful evidence within that scope, not a census of every AI answer.

A comparison can also change for reasons outside your edit: user demand, query mix, seasonality, competing pages, content freshness, model changes, partner refresh cycles, tracking changes, and sales-cycle timing. Keep an annotation for material changes and resist explaining a movement with one cause when several changed together.

A simple percent change is (comparison value minus baseline value) divided by baseline value, multiplied by 100. When the baseline is zero, report the absolute difference and write 'percentage change not defined' rather than inventing a percentage. For small counts, show both the raw numbers and the percentage.

  • Do not treat an unavailable report, missing crawl date, or unclassified visit as zero.
  • Do not combine impressions, citations, sessions, and outcomes into a single score.
  • Do not infer a provider-wide trend from a small hand-run prompt sample.
  • Do not claim causation from a before-and-after comparison without a design that can rule out alternatives.

Choose repeatable questions from real demand

Start with Search Console queries, customer questions, and sales objections for the page you are measuring. Cyrus Shepard’s Zyppy Signal fan-out framework suggests beginning with a topic where your page already has search visibility, then reviewing related questions before measuring again. This is attributed practitioner guidance for choosing useful coverage, not a confirmed ranking rule.

Keep the same question set for comparisons and save the exact prompt, engine and model, collection date, locale, retrieval setup, answer, and cited URLs. Fan-out queries and answers can vary between runs. A sample can reveal an inaccurate brand description or missing evidence, but it cannot establish market-wide citation share.

Choose the next action after the baseline

Use the worksheet to find the first evidence-backed gap. If indexing is unresolved, investigate access and canonical signals before interpreting low AI visibility. If the page is eligible but platform reports show little activity, review the page and the questions it serves without promising a citation. If visibility is reported but visits or qualified outcomes are absent, check the landing-page experience, referral attribution, and the offer before publishing more content.

This guide focuses on measurement, not the order for fixing every SEO or GEO problem. Use the related guides to choose an action after the baseline or interpret a crawl-date observation.

Copy the repeatable worksheet

Choose one page and two equal-length periods. Use the same property, timezone, geography, device filters, and outcome definition in both. Fill in what each source actually reports; use 'not available' or 'unknown' where evidence is missing.

Scope and comparison
1. Indexing status

A technical eligibility check, not a visibility count.

Source: Search Console URL Inspection

Verdict, inspection date, canonical, and crawl date if shown

Keep status and dates as text. An indexed verdict does not establish an impression or citation.

2. Supported AI impressions and citations

Keep each provider's report, feature, and unit in its own row.

Source: Google Search Console and Bing Webmaster Tools

Google Search generative AI impressions (impressions)

Bing AI Performance citations (citations)

Bing cited pages and grounding-query notes (notes)

Google impressions and Bing citations have different definitions and are not additive or directly comparable.

3. Referral visits

Count sessions identified by the same source rule in both periods.

Source: Analytics, filtered to this landing page

AI-attributed sessions (sessions)

Source and channel notes (notes)

A reported citation is not a visit. Unattributed sessions remain unknown.

4. Business outcomes

Choose one verified event and preserve its definition.

Source: Analytics key event, CRM, booking, or commerce system

Outcome count and value (count or amount)

Attribution window and rule (notes)

A before-and-after change does not prove that an AI impression or citation caused the outcome.

Worked example: read each change in its own unit

Hypothetical values created only to demonstrate the arithmetic. These are not AEO Improvement, customer, or research results.

Suppose the same page is indexed in two equal four-week periods. The provider reports and the site's analytics show the following:

  • Indexing: Indexed to Indexed (No status change). The eligibility check stayed the same. It does not establish AI visibility.
  • Google generative AI impressions: 80 to 100 (+20, or +25%). The Google report's impression count rose for its reported scope.
  • Bing AI citations: 4 to 6 (+2, or +50%). Bing reported two more citations in this small sample. It is not a 50% increase in visits.
  • Identified AI referral sessions: 8 to 10 (+2, or +25%). Analytics attributed two more sessions under the chosen source rule.
  • Qualified leads: 1 to 1 (No change). The selected business outcome did not increase in this example.

The defensible summary is that the reported counts changed while qualified leads did not. The sample does not show that citations generated the sessions, that sessions generated the lead, or that an edit caused any movement. Check the raw counts, report availability, attribution rules, and other changes before choosing a next step.

Sources

Claims about provider reporting are linked to current official documentation. Values in the worked example are hypothetical.

Start with a measurable page

Choose the page, source, period, and outcome you can verify before making changes.

Run a free page audit