In short. A correlation with citations provides an association; a causal claim needs evidence that isolates the effect of changing the page feature.
Definition: No. A correlation describes an observed relationship between a page feature and a citation outcome. A causal claim concerns the effect of changing that feature, which requires evidence that separates the change from other explanations.
Key facts
- The Citation Selection to Citation Absorption study treats page-feature associations as descriptive evidence and reserves causal prescriptions for intervention experiments.
- Citead's GEO score paper reports that quotation, statistics and cited-source edits produce no positive transfer of the historical results across ten tested engine families.
- The NIST guide to completely randomized designs assigns experimental conditions randomly and uses repeated runs to compare responses.
What does observational evidence establish?
The April 2026 Citation Selection to Citation Absorption paper distinguishes citation selection, a page's inclusion as a source, from citation absorption, its measured contribution to the answer. Its influence score combines observable text relationships and citation-position features.
The study associates more structured pages with greater measured absorption. Headings may also accompany better editorial quality. Confounding occurs when another variable offers an alternative explanation for an association; the authors explicitly identify editorial quality as a possible shared explanation.
The paper also warns against explaining its influence score with components already used to calculate it. Such a relationship follows partly from the measurement formula itself.
What changes when researchers intervene?
An intervention deliberately changes a feature and compares outcomes under specified conditions. In the September 2026 Citead paper, each query supplies five source texts. The target source receives a quotation edit or neutral filler matched for length; unchanged sources provide a baseline. This comparison separates the quotation edit from the effect of adding text.
The outcome is a source's share of position-weighted words in sentences citing it. The paper's quotation, statistics and cited-source interventions fail to reproduce the historical positive effects. These findings concern supplied texts under a fixed citation prompt, with retrieval and open-web competition outside the experiment.
The paper names this released report for inspecting the pooled intervention estimates:
modern_anchors_report.json
What should a credible test specify?
The NIST experimental-design glossary defines a factor as an input manipulated by an investigator, a response as an output, and replication as repeating a treatment combination. Applied to a page-edit study, these definitions give a practical reading checklist:
- The factor identifies the exact edit being tested, such as adding headings.
- The response identifies whether the study measures citation selection or an absorption score.
- Randomization specifies how conditions and their run order are assigned.
- Replication repeats conditions so that random variation can be estimated.
This checklist is an application of NIST's design principles to the citation question. A report should make the comparison explicit enough to identify which change its result concerns.
Related terms
- Citation selection and absorption study
- The GEO score paper (arXiv 2609.07559)
- AI answer reproducibility
- AI citation
FAQ
Can an observational study guide a website edit?
Yes. The citation absorption paper proposes matched rewrites as a follow-up experiment. You can treat an associated feature as a hypothesis to test.
Does a higher GEO score establish a citation effect?
No. Citead's paper positions its query-independent score as a quality filter. Raising that score alone leaves the causal effect on citations unresolved.
Sources
- arXiv, "From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms", 29 April 2026
- arXiv, "Scoring Without the Engine: Validating a Deterministic, Manipulation-Resistant Content Score for Generative Engines, End to End", 7 September 2026
- NIST/SEMATECH, "Completely randomized designs", undated
- NIST/SEMATECH, "A Glossary of DOE Terminology", undated
Check how a page reads to answer engines. Open the GEO score analyzer