An association in pooled data can disappear or reverse within groups because the mix of observations differs.
The overall rate can conceal group differences
Invented example: method A succeeds on 9 of 10 easy tasks and 30 of 100 hard tasks. Method B succeeds on 80 of 100 easy tasks and 2 of 10 hard tasks. A performs better within both groups, yet its overall rate, 39/110, is lower than B’s 82/110 because A handles far more hard tasks.
Why it matters for GEO evidence
Large sites may receive more AI citations and also use FAQs more often. A positive association across sites may disappear when pages are compared within the same site. Neither a causal benefit of FAQs nor their universal uselessness follows from these observations.
Grouping needs a causal rationale
Choose groups based on the research question and the role of each variable. A changed association prompts a review of the explanation; it does not automatically establish which analysis is causal. The brand-ranking article (Traditional Chinese) illustrates the distinction between across-site and within-site evidence.
Sources
Definitions and source documents checked on 2026-11-07. Examples marked as illustrative are not measured results.