Does GEO actually work? What the peer-reviewed evidence says

AI assistants now answer questions and cite a few sources, and a whole industry promises to get you "recommended by AI." Most of it claims a certainty the published research does not have. Here is what the peer-reviewed evidence actually shows.

The science is split on whether GEO tactics even work

Two papers, both at top venues, reach opposite conclusions:

• GEO (Aggarwal et al., KDD 2024) found that content tactics like adding statistics, quotations and citations can lift a page's visibility by up to 40% in generated answers. But this was a single-actor, idealized setting: one page optimizing in isolation.
• C-SEO Bench (Puerto et al., NeurIPS 2025) tested the same idea across many domains and, crucially, in realistic settings where many sites compete. Its finding: these methods are largely ineffective and frequently have a negative impact on ranking, with close to zero gain. And any gains are zero-sum: as more sites adopt a tactic, the advantage disappears.

Two peer-reviewed papers, opposite results. The honest conclusion is the one nobody selling "guaranteed GEO" wants to say: no one can promise AI will recommend you. The effect of any tactic is uncertain, and in competitive reality often nil.

What the evidence does agree on

• Being retrievable and authoritative beats content tricks. C-SEO Bench found traditional strategies (making your source rank and get retrieved) far more effective than conversational-SEO formatting.
• Substance beats packaging. Where content helps, it is evidence density (statistics, comparisons, definitions, original data), not formatting. FAQ / Q&A blocks on their own don't help (recent analysis), and keyword stuffing didn't help even in the optimistic GEO study.
• Brand authority and mentions across the web correlate with citation far more than backlinks do (independent industry studies).

So if a tool is selling you schema markup and FAQ blocks as "GEO," it is selling certainty the research does not support.

Even being cited is not the whole story

A peer-reviewed Nature Communications study (SourceCheckup, 2025) found that between 50% and 90% of LLM responses are not fully supported by the sources they cite; even GPT-4o with web search leaves roughly a third of its statements unsupported. So being cited is not enough; being represented accurately matters too, and any honest measurement has to account for how noisy these systems are.

Why you measure instead of guess

• Engines cite almost different universes of sources. The same brand can be visible on Gemini and nearly invisible on Perplexity.
• AI citations are stochastic. Ask the same question twice and you get different sources; a single check is noise. Recent work recommends repeated sampling with confidence intervals to get a reliable estimate, which is exactly how we measure.

That is the whole point of Citedar: because the science cannot promise a tactic will work, we measure your real citation rate across ChatGPT, Gemini and Perplexity, in English and Hindi, with statistical rigor, and we only call a recommendation "validated" once a change has actually moved your numbers.

Sources

Peer-reviewed: GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024) · C-SEO Bench: Does Conversational SEO Work? (Puerto et al., NeurIPS 2025) · SourceCheckup, on how well LLMs support their citations (Nature Communications, 2025). Supporting preprints (2026, not yet peer-reviewed): citation selection vs absorption; quantifying uncertainty in AI visibility. Plus independent industry citation studies. Specific preprint and industry figures are directional; the peer-reviewed work and cross-source agreement are what we rely on.