The Confound Problem in Observational GEO Research: Why 21-Site Comparisons Cannot Support Causal Claims

1. Introduction

The emerging field of Generative Engine Optimisation (GEO) faces a fundamental methodological challenge: much of the available evidence about what drives AI citation comes from observational comparisons between sites that differ on dozens of variables simultaneously. When a cited site has more H2 headings, more external links, higher entity density, AND is five years older than an uncited site, attributing the citation difference to any single factor is logically invalid.

This paper makes the confound problem explicit by cataloguing every comparison in the Category C paper series (SIGI-2026-036 through SIGI-2026-049) that fails Gate 2 of the Logic-First Research Methodology, explaining why it fails, and establishing the inferential boundaries for what can and cannot be concluded from the data.

We make this analysis not to invalidate the observational work — which has substantial value for hypothesis generation and descriptive characterisation — but to prevent over-interpretation. The GEO industry is young enough that establishing rigorous inferential standards now will prevent the accumulation of unfounded causal claims that plagued early SEO research.

2. The Confound Structure

2.1 Variables That Co-Vary with Citation Status

Every row in this table co-varies with citation status. No statistical technique can separate these effects from observational data alone.

2.2 The 22 Confounded Comparisons

3. What CAN Be Concluded

3.1 Necessary/Sufficient Tests (Gate 3)

Unlike population comparisons, necessary/sufficient tests require only single counterexamples and are NOT invalidated by confounding. The following conclusions survive Gate 2:

• Content volume is NOT sufficient for citation (counterexample: high-volume sites with zero citations; SIGI-2026-043)
• Schema markup is NOT sufficient for citation (counterexample: highest-schema site scores zero; SIGI-2026-038)
• Schema markup is NOT necessary for citation (counterexample: low-schema sites achieve citation; SIGI-2026-038)
• Image count is NOT necessary for citation (counterexample: zero-image site achieves score 7; SIGI-2026-046)
• Image count is NOT sufficient for citation (counterexample: moderate-image sites score zero; SIGI-2026-046)

3.2 Descriptive Statistics

All means, medians, ranges, distributions, and taxonomic classifications are valid as descriptions of this specific sample. They characterise what was observed, without claiming generalisability or causation.

3.3 Hypothesis Generation

Every confounded comparison generates a testable hypothesis. The 22 comparisons above translate directly into 22 controlled experiments that could be designed to isolate each variable:

4. Why This Matters for the GEO Industry

The GEO discipline is at a critical juncture. Early SEO research suffered from decades of confounded observational claims being treated as established facts (“keyword density drives ranking,” “more backlinks always improves position”). Many of these claims were eventually falsified by controlled testing or algorithm changes, but not before substantial industry resources were misallocated.

GEO has the opportunity to avoid this pattern by establishing rigorous inferential standards from the outset. This paper contributes to that goal by making the confound structure of the most commonly available data type (observational website comparisons) explicit and actionable.

When a GEO practitioner reads that “cited sites have 2.3x more H2 headings,” they should immediately recognise this as a confounded observation that does not justify adding more H2 headings. The appropriate response is to generate a testable hypothesis and seek controlled evidence before allocating resources.

5. Limitations

• This paper critiques its own dataset: The confound analysis applies to the specific 21-site dataset used across Category C. Different datasets with different composition may have different confound structures.
• Not all confounds are equal: Some confounds (e.g., domain age) are likely more impactful than others (e.g., hreflang count). This paper treats all confounds equally, which may overstate the limitation for some comparisons.
• Controlled alternatives exist: The Category A papers (SIGI-2026-001 through SIGI-2026-020) provide controlled evidence for many of the variables examined observationally in Category C.

6. Conclusions

All observational comparisons in the 21-site competitive intelligence dataset are maximally confounded. We catalogue 22 specific confounded comparisons to prevent over-interpretation and guide future controlled experiments. What can be stated: necessary/sufficient tests via counterexamples, descriptive statistics, and hypothesis generation. What cannot be stated: causal claims about any individual variable.

The GEO industry should treat observational website comparisons as hypothesis generators, not evidence generators. Causal claims require controlled experimentation with single-variable isolation, as demonstrated in the Category A paper series.

Confidence: HIGH for the methodological critique. This is a meta-analytical paper about research limitations, and the confound structure is demonstrably present in the data.