The SEO Meta Study: A Systematic Review of 9,249 Documents on SEO, AEO and GEO

What this is. The SEO Meta Study is a systematic review with narrative synthesis of published evidence on search engine optimisation (SEO), answer engine optimisation (AEO) and generative engine optimisation (GEO). It screens a corpus of 9,249 documents, filtered from a crawl of over 30,000, and reports 182 tactics that carry a measured outcome and an identifiable source. 552 primary sources across 268 publications are cited, each with a live link. The report is 170 pages and free to download.

Method. The prefilter is deliberately permissive and passes 98.4% of the corpus, because a keyword filter scored against an author-verified reference set of 90 documents measured precision 0.32 and recall 0.48, and would have discarded half of the real case studies in that set. Screening runs as a language model prepass measured against the same reference set. Every document that reaches extraction is read twice, blind, by models from two different providers, with no cross-provider fallback. A finding counts only when its quote appears literally in the source document and its numeric value is located inside that quote, so every printed figure is the figure its publisher printed. Interventions are assigned to a frozen controlled taxonomy of 1,595 variants across 28 families rather than to embedding clusters, and the largest variant holds 1.1% of assigned findings. One published document counts as one sample, so two extraction passes over the same document, or two segments of one A/B test, are not two samples. Labels naming several simultaneous changes are refused entry to the taxonomy, because one measured value for three changes supports none of the three.

Main findings. Of 769 tactics scored, 602 have no controlled test behind them, 138 rest on exactly one, 23 on two, and 6 on three or more. Measured on the corpus screen, 81.9% of published outcomes are positive and 3.8% are negative; on the 90 document reference set the same skew reads 64% and 8%, and neither frame should be quoted as the other. Of 118 credited controlled measurements, 65 come from a single publisher. 92 of the 182 entries stand on a study that changed several things at once while measuring one number, and are marked as such. The median reported gain across the 87 entries whose headline metric is organic sessions is +36%.

What this is not. This is not a meta-analysis and is not presented as one. Pooling effects statistically requires a variance per study, and 4 of the 25 usable studies in the reference set report one, so no average per tactic is printed anywhere in the review. Each entry reports the median across samples with the sample count beside it.

Limitations. The corpus is what the open web publishes, so the review inherits publication bias, and the 3.8% negative rate is almost certainly not the failure rate of SEO work. The reference set was verified by the author rather than coded independently by two people. The academic portion largely stands on abstracts, because older arXiv entries expose nothing else. A document reporting several genuinely independent tests is counted once, which is conservative. Non-English documents were out of scope. Evidence in this field has a short shelf life: of 793 cited source records, 523 are dated 2025 or 2026.

Availability. Free, with no registration and no email required, at https://ovisy.hr/blog/seo-meta-study.