What can GEO measure? An evidence model for AI-search visibility

A team asking whether pages are eligible needs crawl and index evidence. A team asking whether Google shows those pages in generative features needs Search Console exposure data. A team asking whether an AI answer understands NobleJackal needs captured answers and a factual rubric. A sales question requires intake and CRM evidence. No single chart answers all four.

Write the decision before collecting the metric: fix access, expand a source, correct entity facts, improve buyer relevance or change the offer. This prevents an available number from becoming important merely because the platform happens to display it.

A rate requires the number of eligible opportunities. Ten citations from ten runs and ten citations from a thousand runs tell different stories. So do fifty indexed pages from sixty submitted canonicals and fifty from six hundred. Publish the numerator, denominator, exclusions and missing observations together.

Platform totals may use property-level aggregation while tables use page-level aggregation. Do not force them to reconcile by editing one total. Retain the provider's rule and compare the same dimension in each period.

A global total can grow while German buyer-intent exposure disappears or Turkish informational traffic dominates every other locale. For a six-language programme, the minimum useful cell is page × language × country × device or surface × date; controlled answers add platform and intent family.

Only aggregate after the underlying cells remain inspectable. Low-volume groups can be reported over a longer window, but they should not be hidden inside English or blended into a fabricated Europe-wide rank.

Record the exact release, affected URLs and first eligible observation date. Search and answer systems need time to crawl and refresh, while preliminary reports can change. A same-day lift cannot automatically be assigned to the deployment that happened that morning.

Compare equivalent windows and keep seasonality, reporting delays, product changes and unrelated campaigns in view. If the prompt panel, geography or account state changes, begin a new cohort rather than rewriting the baseline.

A citation counter cannot show whether the link supported the central claim, a minor fact or merely appeared in a source tray. Resolve redirects, store the canonical target and compare the cited passage with the generated statement. A brand name may come from a third-party page rather than the brand's own domain.

For portrayal, use material facts: organisation, founder, expertise, location, service scope, publications, pricing and limits. A fluent answer can still be wrong. Mark uncertainty and partial support instead of treating polished prose as accuracy.

Search Console describes activity before the click; Analytics describes behaviour after a measurable visit. Their numbers need not match because their systems, canonical handling, time zones, consent and session definitions differ. Use landing page, country, device and period to look for compatible patterns rather than forcing equality.

A qualified brief or sale is strongest when the intake record, landing path and source evidence agree. Even then, call the result attributed or assisted under the declared rule. Long buying journeys, direct return visits and offline conversations mean GEO may be one influence among several.