The 89.0 percent for Google AI Overviews checks each claim against all pages the Overview cites, and SourceCheckup's 75.7 percent counts a statement as supported if at least one source given in the same response supports it. Both are claim-level rates of support somewhere in the reference set: more lenient than checking the attached citation, and an upper bound on the share of claims supported by their own attached citation. They are not bounds on the share of citations that support their sentence, which has a different denominator.

Each parent states its unit in its Statement. The step is set logic over claims: if the attached citation supports the claim, then some cited page does, but not the reverse. It says nothing about a rate counted over citations: a claim with three supporting citations and a claim with one non-supporting citation give 50 percent of claims and 75 percent of citations. It does not estimate how far the attached-citation rate lies below; neither paper prints it.

A per-link audit on either dataset finds the share of claims supported by their attached citation equal to the any-source rate, which would make the bound tight. Query: per-link labels in the AI Overviews release.

Google's AI Overviews get their citations right 89 percent of the time. The 89 percent is claims supported by any of the cited pages; how many links support their own sentence was not counted.