Preserve the answer evidence before interpreting it
Store the exact buyer question, provider, model identifier where available, timestamp, full answer text, returned source URLs, target citation state and provider result status. Only successful responses containing answer text belong in a claim comparison. Failed and unavailable responses must remain excluded rather than becoming omissions or negative evidence.
Keep a versioned question portfolio and provider set. Comparing rewritten questions or a different provider mix can create artificial movement. Preserve non-comparable runs with an exclusion reason and begin a new baseline when the denominator changes.
Segment exact statements, not vague themes
Split each verified answer into readable sentences while retaining provider and response-level citation context. Extract numeric values, dates, units and fact type. Useful classes include quantitative facts, dated facts, comparisons, methods, limitations and general factual statements. The exact sentence should always remain visible beneath any label.
Cluster semantically related statements across providers with a disclosed language-aware method. Token overlap is transparent and repeatable, but it is not human-level understanding. A cluster becomes cross-provider consensus when multiple providers express a sufficiently similar fact. A statement appearing in only one provider remains provider-specific rather than automatically being called wrong.
Treat numeric disagreement as a verification candidate
When related statements contain different extracted values, create a review item containing every sentence, value set, provider, response-level citation URL and matched owned claim. Do not decide which value is correct from provider frequency. The difference may reflect year, geography, population, unit, scenario, method or a genuinely stale record.
Agreement does not prove truth, and disagreement does not prove error. The primary evidence and declared scope decide what the public owner should say.
Define omission narrowly
Two omission views are useful. A provider omission occurs when a statement cluster appears across several providers but is absent from another response in the same fixed test. An owned-evidence omission occurs when a material claim on the mapped canonical page is not reflected in any verified answer. Neither state proves suppression, retrieval failure or provider fault.
Owned claims should come from the measured public-page evidence graph. Map them to the question’s owning URL and compare exact language. Report how many owned claims were available and reflected for each provider. If no owner is mapped, show that missing connection rather than inventing a completeness score.
Keep citation scope honest
Many provider interfaces return a list of sources for the response without a durable statement-to-source mapping. Label these URLs response-level. Do not claim that one returned URL supports every sentence. Where a provider supplies exact inline attribution, preserve that stronger relationship separately.
An uncited response is not automatically inaccurate. It is a review priority when the answer contains material numbers, comparisons or brand claims that professionals may repeat. The appropriate action is to inspect the answer, primary evidence and provider citations before updating the owned page.
Turn claim patterns into governed work
- Research and legal: resolve numeric disagreements against current primary evidence.
- Editorial: publish explicit period, geography, sample, unit and limitations where multiple values are legitimately different.
- GEO strategy: strengthen an omitted owned fact only when it is accurate, decision-relevant and properly sourced.
- Web operations: preserve the canonical owner and make its exact answer passage retrievable.
- Measurement: rerun the unchanged question and compare exact wording, values, omissions and source URLs separately.
Export enough evidence for independent review
Each row should contain buyer question, query state, claim state, fact type, provider coverage, included and omitted providers, exact statements, extracted value sets, response-level source URLs, matched owned claims and match score, owner URL, movement, action, verification and evidence class. A professional should be able to challenge the conclusion without reverse-engineering a composite score.