generative search now answers questions about companies directly and cites sources, and nobody checks whether those sources actually support what it says.
In 2025 a Google AI Overview told searchers a Minnesota solar installer was being sued by the state Attorney General. It was not. The answer cited four sources and none of them contained the claim. The company reported losing a $150,000 contract and sued Google; that litigation is ongoing.
Google has since corrected that answer, which is the point — generative answers change underneath you, so this has to be measured continuously rather than once.
There is a whole category of AI-visibility tools and they all measure the same thing — is the brand mentioned, in which prompts, with what sentiment.
None of them check whether the sources cited beside a sentence actually contain that sentence. The missing property is attribution integrity.
Target entity is registered, aliases and collision-set neighbours are indexed, and the search budget is established.
Deterministic probe families are compiled across identity, adverse, commercial, qualification, and operational queries.
Probes are dispatched across a geographic locale grid, capturing append-only raw engine observations and suppression records.
Generative answers are parsed into atomic factual assertions mapped to their containing block text and cited reference indices.
Every atomic assertion is tested against the union of its block citations via SerpApi snippet evaluation and verified page retrieval.
Observed assertions are clustered across locales and engines to compute attribution integrity, consistency, and overall grounds scores.
UNVERIFIABLE is never counted as a defect, because an unreadable source must never become an accusation.