GEO replications and critiques

The main independent retest of the 2023 GEO methods is C-SEO Bench (Puerto et al., arXiv 2506.11097, NeurIPS 2025 Datasets & Benchmarks). Across six domains and four models it found that most content-rewriting methods did not improve, and often worsened, a document’s citation rank, while a document’s position in the retrieved context had a far larger effect. It qualifies rather than refutes the GEO paper, because the two studies use different metrics.

What did C-SEO Bench test?

• Tasks: product recommendation (Retail, Video Games, Books) and question answering (Web, News, Debate): more than 1,900 queries and 16,000 documents.^[1]
• Methods: the GEO paper’s eight content methods (Authoritative, Statistics, Citations, Fluency, Unique Words, Technical Terms, Simple Language, Quotes) plus two new ones, Content Improvement and an llms.txt-inspired “LLM Guidance.”^[1]
• Models: gpt-4o-mini, claude-3-5-haiku, o3 and o4-mini.^[1]
• Metric: improvement in citation rank (being cited earlier), with Wilcoxon tests and Holm–Bonferroni correction.^[1]

What did it find?

• The authors report that out of 54 cases only three showed statistically significant gains: LLM Guidance in Retail and Video Games, and Content Improvement in Retail, none of them original GEO methods. “We found no method effective for the question answering task.”^[1]
• Many methods significantly reduced rank: “the Statistics method decreases rankings in 19 out of 24 evaluated settings.”^[1]
• Placing the document first in the model’s context gave much larger gains (e.g. +2.77 positions in Retail on gpt-4o-mini, versus +0.36 for the best rewrite).^[1]
• Gains fall as more competitors adopt a method: a “congested and zero-sum” game.^[1]

Does this contradict the GEO paper?

Partly. The C-SEO authors attribute the difference to metrics: the GEO paper’s headline results use word count, which “does not measure the LLM preference,” and they note the GEO paper’s position-adjusted results show smaller effects. They conclude “the results of both papers on LLM preferences do not contradict each other.”^[1] Practically: rewriting content to add statistics or quotes may make the answer use more of your words, but is not shown to get you cited earlier.

What other research qualifies GEO claims?

• Citation accuracy: Liu, Zhang and Liang (2023) found only 51.5% of generated sentences were fully supported by citations across four generative search engines, and 74.5% of citations supported their sentence.^[2] Citation is not the same as faithful representation.
• Brand signals: Ahrefs’ 75,000-brand correlation study (2025) associated AI Overview brand visibility more with web mentions than with links, without claiming causation.^[3]
• Vendor guidance: Google says rewriting content “just for AI systems” is unnecessary.^[4]

What should practitioners take from the replications?

1. Treat retrieval (classic SEO) as the main lever for citations.
2. Do not mass-rewrite pages to insert statistics or quotations for AI; add real sourced data where it helps readers.
3. Expect any tactic’s advantage to shrink as competitors copy it.
4. Measure with citation-based metrics on your own prompt set.

Frequently asked questions

See also

• The 2023 GEO paper
• Generative engine optimization
• Studies & papers
• GEO SEO
• GEO examples

References

Sources accessed October 14, 2026 unless another date is given. See all sources.

1. “C-SEO Bench: Does Conversational SEO Work?”. arXiv (Puerto, Gubri, Green, Oh, Yun). June 2025; NeurIPS 2025 Datasets & Benchmarks.
2. “Evaluating Verifiability in Generative Search Engines”. arXiv (Liu, Zhang, Liang; Stanford). April 2023; Findings of EMNLP 2023.
3. “An Analysis of AI Overview Brand Visibility Factors (75K Brands Studied)”. Ahrefs. May 26, 2025.
4. “Optimizing your website for generative AI features on Google Search”. Google Search Central. accessed October 14, 2026.