Why do GEO studies disagree with each other?

Most GEO studies that look like they conflict are measuring different things. A study that moves a document to the top of a model's context window is not testing the same thing as a study that tracks citations in live Google AI Overviews. Both may be sound. They just answer different questions.

Five causes explain most of the disagreement:

• Parsing and definitions. One team counts URL matches, another counts domain matches, and a third never defines its metric.
• Lab engines versus live engines. Many experiments use an engine the authors built.
• Content already in context. The effect is measured after retrieval has already happened.
• Vendor incentives. The study's author often sells the fix the study supports.
• Brand-size confounding. Big brands get cited more, and few studies can control for it.

The July 2026 critical survey of 45 GEO studies puts the result plainly: already-retrieved content can causally alter its citation or use, but no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability. That gap is where most of the conflict lives.

Take the question "how many AI Overview citations come from pages ranking in Google's top 10?"

Ahrefs found in March 2026 that 37.9% of URLs cited in AI Overviews also appeared in the first 10 results (37.1% counting organic results only). BrightEdge reported in September 2025 that only 16.7% of citations came from the top 10, while also claiming 54.5% overlap. BrightEdge published no sample size and did not define "overlap."

Those two figures do not corroborate each other, and they cannot simply be averaged. One method is wrong, or they measure different things.

There is a second trap inside the Ahrefs data. An earlier Ahrefs study put the figure near 76%. It is tempting to read 76% to 38% as a collapse. But Ahrefs changed its parsing method between the two studies, so the endpoints come from different instruments. The 38% figure stands. The decline is not a measured trend. Our digest of the Ahrefs top-10 overlap study covers the method in more detail.

URL-level matching adds another distortion. A large site cited through a URL that does not itself rank gets counted as a miss, which penalises big sites mechanically.

A large share of GEO experiments never call a live search engine. They hand the model a fixed set of documents and measure which one gets cited.

The original GEO paper (KDD 2024) reported a 30 to 40% visibility band across three methods. It used a two-step engine the authors built, with sources supplied in context. Its one Perplexity test uploaded source text as files, because Perplexity did not let users specify source URLs. That is a Stage 3 finding (which already-retrieved document gets used), often sold as a Stage 1 or 2 outcome (whether your page gets found at all).

Better-designed lab studies reach the opposite conclusion. C-SEO Bench (NeurIPS 2025) found most conversational SEO methods largely ineffective and frequently harmful to document ranking. Its "traditional SEO" arm worked by moving the target document to position one in the context, which no publisher can do. The 252,000-trial factorial used two candidate documents per trial on a synthetic corpus, which likely overstates position effects.

So a lab "40% lift" and a lab "no effect or negative" can both be true inside their setups. Neither tells you what happens on a live engine. See lab GEO studies versus live engines for what transfers.

Much GEO research comes from companies that sell tools or services. That does not make the data wrong, but it shapes which questions get asked and how results get framed. The 252,000-trial factorial was written by an industry team and ships a GEO checklist. A review-profile study tied to a Trustpilot white paper never defined "citation rate." We sell AI visibility audits too, which is why we cite primary sources and flag limits.

Brand size is the bigger problem. Semrush's brands visible across all four surfaces it tracked included YouTube, Google, Reddit and Amazon: the finding is brand size. A 2025 study of AI search arena data found the top 20 news sources accounted for 67.3% of citations for OpenAI models. Studies that try to control for this use metrics like Domain Authority or Domain Rating, which are built from backlinks and traffic, so they are proxies for the confounder itself. Our page on correlation versus causation in AI search studies explains why that does not remove the bias.

Even careful work has boundaries. Ahrefs' matched-control schema study tracked 1,885 pages and found AI Overview citations fell 4.6% after schema was added. Every treated page already had 100+ AI Overview citations. The study cannot say what happens to pages with none.

Run any study through these questions before changing your roadmap:

1. What exactly was measured? Is "citation," "overlap" or "visibility" defined, with a formula?
2. Which engine, and was it live? A lab engine with documents in context is not ChatGPT or AI Overviews.
3. Was retrieval on? Published estimates for how often ChatGPT searches the web range from about 18% to 80.5%, depending on who measures. An answer with no search reflects training data, not your page.
4. Is there a control group? A study of only cited pages cannot tell you what makes pages get cited.
5. Did the method change between data points? If yes, a trend is not a trend.
6. Were brands of different sizes compared? If not, the result may be brand size with a GEO label.
7. Who funded it, and what do they sell?
8. When was it run? Engines change. Treat anything older than a model generation with caution.

Then act on what holds up across sources. As of October 2026, the strongest evidence sits with access: server-rendered content, per-bot robots.txt permission, firewall reachability and index eligibility. Our AI SEO best practices, sorted by evidence lays these out. Because engines differ, check whether results on one engine carry to others before generalising. If you want a baseline for your own brand, the free checker on this site shows whether AI engines name you for three buyer questions.