What does the research establish—and what remains uncertain?

Transparent sampling and evidence retention are defensible requirements; a universal winning GEO formula is not established by the sources reviewed. AMEC’s GEO Principles call for governed queries, documented collection conditions, repeated testing, saved outputs and separation of visibility from outcomes. NIST’s generative AI profile separately recommends documenting measurement bias and variance and assessing third-party risks. Neither organisation endorses this scorecard’s numerical weights. (amecorg.com)

Recent research reinforces the need for caution. A March 2026 preprint examined citation variability across three consumer-product topics, but explicitly limited its generalisability to B2B settings. A May 2026 preprint found substantial recommendation changes under prompt rewording, while acknowledging its single-day measurement window. These are reasons to test reliability locally—not evidence for a universal minimum prompt count. (arxiv.org)

Which 2026 changes should alter the procurement brief?

Ask suppliers to reconcile their monitoring with current first-party reporting—not repeat outdated claims that no such reporting exists.

• Google: its current help documentation describes a dedicated Generative AI performance report covering impressions in AI Overviews and AI Mode, with worldwide rollout stated as 31 August 2026. A report may be absent when a property has insufficient impressions; that is not proof of zero visibility. Older documentation still describes AI-feature traffic within the overall Web report, so record which documentation and property view underpin the supplier’s analysis. (support.google.com)
• Microsoft: Bing’s June 2026 announcement added preview capabilities for Intents, Topics, Citation Share and Compare. Microsoft explicitly describes Citation Share as observational: it does not expose competitor domains or represent traffic share. (blogs.bing.com)
• Google technical controls: current documentation also describes a Search generative AI inclusion control in Search Console. A technical audit should inspect the property’s actual setting rather than rely solely on an older crawlability checklist. (support.google.com)

Procurement implication: require separate reporting for first-party impressions or citations, supplier-generated prompt observations and website outcomes. Do not combine these different denominators into a supposedly universal market-share number.