AI Search

What 45 GEO Studies Actually Prove — and What They Don't

Generative engine optimisation has a playbook — add statistics, quote experts, use clear structure, publish lists — and a growing stack of studies that claim to prove it works. A critical survey published in July 2026 reviewed 45 of those studies and found that the tactical playbook mostly does not hold up under scrutiny. A November update from Trakkr adds a second problem: many of the statistics AI answers cite do not trace back to a study that says the same thing.

The survey of 45 studies

The critical survey, by Martinez and published on arXiv in July 2026 as arXiv:2607.14035, reviewed 45 GEO studies published between November 2023 and July 2026. The survey's conclusion was blunt: the tactical playbook — the familiar list of things to add to a page to get cited by AI — mostly does not hold up when the underlying studies are examined closely.

That matters because the playbook has become industry consensus, repeated in blog posts, conference talks, and agency decks as though it were settled science. The survey suggests much of it rests on weak foundations: small samples, synthetic benchmarks, results that do not replicate, and findings overstated in the telling. A tactic that worked once, on one set of test queries, is not a tactic that works in general — and the field has too often treated it as though it were.

None of this means GEO is a fiction. AI systems do cite sources, and some pages get cited far more than others. What the survey challenges is the confidence with which specific tactics are prescribed: if the evidence for a tactic is thin, following it is a bet, not a best practice, and it should be sold to clients as one.

The Trakkr audit: where the numbers come from

The November 2026 update from Trakkr, dated October 28, 2026, looks at whether the statistics AI answers cite are real. Trakkr traced 1,285 AI-cited statistics across five engines, following each citation trail back to its source. Only 40% reached a study that says the same thing as the cited number. Another 40% stopped at a source that cannot be checked. In 11% of cases there was no study at all, and in 8% the trail reached a study whose number had drifted.

There is one reassuring figure in the audit: 82% of cited pages do contain the number being quoted. So the numbers in AI answers are usually present on the page that gets cited. The problem is what lies behind the page. A statistic can appear on a page and still not be traceable to the study it supposedly comes from — because the page does not link to the study, because the study is behind a wall or has moved, or because the number was copied through several hands and changed along the way. The citation looks supported, and the page checks out, but the evidence chain breaks before it reaches the original research.

For an industry that has spent the year telling clients to add statistics to get cited, that is an uncomfortable finding. The tactic may work — a page with a striking number may well be more citable — but the number itself may not survive scrutiny. And if AI systems are citing statistics that cannot be verified, then the visibility they confer is built on sand.

Why it matters for SEOs and link builders

For SEOs, the two studies together argue for a more sceptical, more empirical approach to GEO. Treat the playbook as a set of hypotheses, not a checklist. Test tactics on your own queries and keep the ones that move your citation numbers. When a vendor study claims a tactic lifts visibility, ask how it was measured, on what queries, and whether the result would survive a different sample. The Martinez survey gives you permission — and a citation — to push back when a tactic is presented as proven and the proof is thin.

For link builders, the Trakkr audit points to an opportunity. If 40% of citation trails stop at a source that cannot be checked, then a checkable page has an edge: a statistic with a direct, working link to the original study, quoted accurately and named. Original research is the strongest version of this — a study you ran yourself, published on your own site, is a citation trail that cannot break. That is the kind of asset that earns both citations and links.

There is also a client-trust angle. Clients are being sold GEO services on the strength of studies that, by this survey's account, mostly do not prove what they claim. Being the agency that explains that honestly — that some tactics are promising, some are unproven, and the honest work is testing — is a better position than being the agency that repeats the playbook and hopes nobody checks the footnotes.

What to watch next

Watch whether the field responds to the survey by improving its methods. The test will be the next wave of GEO studies: larger samples, real queries, pre-registered methods, and results reported with their limits attached. Watch, too, whether AI engines start checking their own citation trails — if a system began verifying that a cited statistic matches its source study, the pages with clean, verifiable evidence would gain, and the 40% of trails that currently cannot be checked would become a liability.

Sources

Sources: Martinez, critical survey of 45 GEO studies, arXiv:2607.14035, July 2026; Trakkr, AI citation audit, October 28, 2026.

Related reading on this site: Yahoo Ends Its Bing Partnership After 16 Years and Bets on Scout, AI Overview Citations: What 15 Studies Agree On, and Don't, AI Overviews Are Replacing the Brand Knowledge Panel, 84% of AI Citations Point to Third-Party Sites, Muck Rack Finds, Cited by AI, Picked by No One: Seer's 362,188-Response Study, and YouTube Takes 24.5% of AI Overview Citations in Ahrefs' November Ranking.

More from Linkbaits

Turn this playbook into links

Linkbaits helps you create content that earns citations and backlinks. Start with the toolkit and build your first linkable asset this week.

Start free trial