CLEARCRAFT INSIGHTS · 2026-10-23
Measuring GEO: keep visibility, citations, completed tasks and enquiries separate
When an owner asks whether a website update brought business through AI, a team may produce search clicks, screenshots of AI answers and enquiry counts. Each can be useful, but they describe different people and events. Putting them on one growth chart does not answer the question.
DIRECT ANSWER
How should GEO be measured?
CLEARCRAFT proposes a four-layer evidence log: keep search visibility, citations in observed answers, completion of defined tasks and verified enquiry outcomes separate. Attach a source, time window and denominator to each measure, and preserve missing evidence as unknown. This is our suggested working method, not an AI platform's scoring standard.
Read the official sources →Start with the decision you need to make this week
Begin with an unresolved business question. Your service page may have changed while customers still ask the same basic questions. Do you need clearer content, a better enquiry flow or better follow-up? Write down one decision before choosing the records that could inform it.
Each layer serves a purpose: visibility records discovery, citations show how an answer uses information, task completion tests the next step, and enquiry outcomes establish whether suitable opportunities arrived. Separate reporting helps prevent an improvement in one measure from hiding a problem elsewhere.
Layer one: preserve what search impressions actually mean
Google says a page must be indexed and eligible for a search snippet to qualify as a supporting link in AI Overviews or AI Mode; eligibility does not guarantee display. Traffic from these AI features is included in overall Search Console Web performance.[G1]
From an authorised report, preserve the property, page, period, filters, impressions and clicks. Label a Web total as a Web total: it cannot establish whether a particular answer cited the brand. A page being accessible now, indexed and observed in search are also different states.
Layer two: check citations one answer at a time
Choose a small set of real buyer questions, separating branded from non-branded questions. Save the exact question, platform and interface, language, configured location, time, visible settings, raw answer and links. If the user's actual location cannot be verified, record it as unknown.
We suggest separate fields for a brand mention, a visible target-domain link and whether that link supports the relevant statement. A company name may appear without a link; a linked page may contain an outdated price. A reviewer should open the source and check its content, date and business identity. An inaccessible source remains unverified.
Make the denominator explicit
These numbers are illustrative, not CLEARCRAFT or client results: 12 planned runs produce 10 reviewable answers and two tool errors. Three of the 10 answers display a target-domain link. Report 'target-domain links in 3/10 reviewable answers; errors in 2/12 attempted runs', with the underlying records.
The resulting 30% describes this answer sample only. It is not overall AI market share or a citation rate for all users. The two errors may affect the result and should remain visible. Define the denominator before testing. If the platform, question set, location settings or proportion of reviewable answers changes, show separate groups and explain the difference.
Layer three: test a defined next step
Write an observable task, such as 'find the service scope, complete a test enquiry and stop before confirming submission'. Define its endpoint first, then record the entry point, required information, errors, assistance and result. When the authorised task ends at preparation, stopping correctly for confirmation meets that endpoint; it does not count as an enquiry received.
Keep human and agent tests separate. Use test data in an approved environment and record where failures occur. Our suggested completion measure is attempts reaching the predefined endpoint divided by attempts started, with aborted, assisted and failed attempts also reported. This measures workflow usability, not search visibility or sales performance.
Layer four: let the business define a qualified enquiry
The person handling enquiries should define qualification first: for example, a matching service area, an in-scope need and usable contact details. Received, qualified, meeting booked and sale completed are different events. Repeated contact should connect to the same internal record.
Classify origins as observed referrals, customer-reported sources, other or unknown. 'I found you through AI' is useful customer testimony, but cannot by itself identify the answer or content change that led to the enquiry. Work from existing approved analytics and business records. Preserve unknowns instead of assigning all direct traffic to AI.
Connect evidence with an index, keeping each time window
A record should contain an ID, layer, observation time and time zone, page or task version, question-set or workflow ID, value and denominator, evidence location, review state, owner and next action. Search reports, answer samples, workflow tests and CRM records can keep their own identifiers. Record a connection only when supporting evidence exists.
Cloudflare Web Analytics provides page-view, visitor and real-user page-performance information.[C1] Where already enabled, use it for experience diagnosis, such as investigating a slow form page. RUM alone cannot establish a citation in an AI answer. Tools may cover different populations, periods and collection methods; their numbers need not share a denominator.
Turn the weekly review into one change you can retest
Check record completeness before choosing a fix. If answers cite old terms, verify and update the source. If visitors find the form but misunderstand required information, clarify the fields. If enquiries are consistently outside the service scope, inspect the audience description. Record the change date, owner and expected acceptance result.
Repeat the same questions or workflow after the change and report observations alongside unresolved issues. Increased impressions or enquiries in the same period may still reflect demand, other channels or platform changes. The log helps a team choose its next action while keeping every conclusion connected to evidence.
OFFICIAL SOURCES · CHECKED 2026-10-23
First-party documents
Revision record
v1.0 · 2026-10-23 · Initial publication in Traditional Chinese and English.
This article describes CLEARCRAFT's proposed recording method. It does not access client analytics, CRM records or private platform data, and reports no actual ranking, citation or commercial results. Distinguish documented product facts from our working recommendations.
Submit verifiable corrections through the editorial and corrections policy.
