Most GEO advice, including ours, has been delivered with more confidence than the evidence supports. That is not a moral failing peculiar to GEO. It is the condition of the whole field, and in September 2026 someone finally counted it.
Krešimir Ćorluka at Ovisy crawled more than 30,000 documents on SEO, AEO and GEO, filtered them to 9,249, and spent two months extracting every measured outcome he could verify against its source. The result is the most useful thing published about our industry this year — and for anyone buying or selling GEO, it should change the conversation.
How much SEO advice is actually tested?
Almost none of it. Of 769 tactics with a measured outcome, 602 had no controlled test behind them at all. Another 138 rested on a single controlled test. Only 6 had three or more.
An uncontrolled before-and-after tells you something moved after you touched a page. It cannot tell you the page is the reason. Seasonality, a competitor's migration, an algorithm update, a mention on a forum — any of them can move the graph, and a single-site before-and-after has no way to separate them out.
GEO has the thinnest evidence of all
This is the finding that should matter most to anyone buying GEO services. The review counted entries by topic, and the two topics the industry talks about most right now came in last.
Seventeen entries between AI visibility and answer formatting, across a review of more than nine thousand documents. The tactics we recommend most often — FAQPage schema, question-led headings, direct-answer paragraphs, entity disambiguation — are mostly reasoned from how retrieval systems work. That reasoning is sound. It has not been measured.
Why published results look so good
This is publication bias, and it is nobody's scandal. Agencies publish the engagements that went well because that is what a case study is for. The consequence is that the published record — including the median published gain of +67% in AI citations across 15 entries — describes the best-case tail, not the typical outcome. Do not budget against it.
What this means if you are buying GEO
Ask every consultant, including us, to label their claims by the kind of evidence behind them. There are four kinds, and they are not interchangeable.
| Kind | What it can tell you | What it cannot |
|---|---|---|
| Controlled test | Whether a single change caused a difference, compared against matched pages that did not change | Whether it will work on a different site, category or engine |
| Before-and-after on a live site | That something moved after a change | That the change is why it moved |
| Mechanism reasoning | Why a tactic should work, given how retrieval systems are built | Whether it does, or how much |
| Anecdote | That it happened once somewhere | Almost anything else |
Half of the strongest-looking entries in the review — 92 of 182 — came from studies that changed several things at once and measured one number. That number belongs to the bundle. It supports no single tactic, which is worth remembering the next time a case study credits a jump in citations to whichever change the author happens to sell.
What we changed
We went back through our own published guidance and removed claims we could not support — timelines we had no data for, audit-frequency figures we had not earned, causal language where we only had a mechanism. Each post now says what kind of evidence it rests on.
And we started building evidence instead of borrowing it. Test 01 is a pre-registered, randomized test of whether FAQPage schema affects AI citation — seven matched pairs across three sites, four engines, a fixed decision rule, and a commitment to publish whatever it finds. The design, seed, prompts and scoring code were published before any data was collected.
It is a small test. Its own pre-registration says it can detect a large effect and cannot reliably detect a small one. That is still more controlled evidence on FAQPage schema than we were able to find anywhere else.
Want your baseline measured properly?
The free snapshot runs five fixed prompts across four engines and tells you plainly what it can and cannot show. No invented timelines.
Get the free snapshot See Test 01How to cite this
Please cite the underlying study for any figure, and keep each number with its denominator.
Ćorluka, K. (2026). The SEO Meta Study: a systematic review of 9,249 documents on SEO, AEO and GEO. Ovisy. https://ovisy.hr/seo-meta-study
Summers, J. C. (2026). The GEO Evidence Problem. Sable Search. https://sablesearch.com/blog/geo-evidence-problem/
Frequently asked questions
Very little. A September 2026 systematic review of 9,249 SEO, AEO and GEO documents by Ovisy found that 602 of 769 scored tactics had no controlled test behind them. Another 138 rested on a single controlled test, and only 6 had three or more.
Not yet in the public record. In the same review, AI visibility produced 10 entries and answer formatting 7, the thinnest evidence of any topic, while content depth produced 36 and title tags 20. Most GEO recommendations are reasoned from how retrieval systems work rather than measured under controlled conditions.
Publication bias. Of 1,700 documents in the review that reported a measured outcome, 1,392 reported an improvement and 65 reported a decline. Agencies and in-house teams publish work that went well and rarely write up work that failed, so the published record overstates how often tactics succeed.
Ask what kind of evidence stands behind each claim: a controlled test, a before-and-after on a live site, reasoning about how retrieval works, or anecdote. Ask how many independent samples it rests on. A consultant who cannot distinguish those categories for their own recommendations is selling certainty that does not exist.
Jason C. Summers
13+ years enterprise technical SEO and digital product. Ran weekly LLM citation monitoring in production at a $1B+ US retailer and, since September 2026, runs pre-registered controlled tests on Sable Search's own properties.
Sources
- Ćorluka, K. (2026). The SEO Meta Study. Ovisy. ovisy.hr/blog/seo-meta-study
- Summers, J. C. (2026). Test 01: Does FAQPage schema affect AI citation? Pre-registration. sablesearch.com/research/faq-schema-test/