Correlation vs Causation in AI Search Studies: How to Read AI SEO Statistics

An observational study takes pages that AI engines already cite and counts what they share. It can tell you that cited pages tend to rank well or carry recent dates. It cannot tell you that ranking or dates caused the citation.

A matched-control study picks pages that made one change. It pairs each with similar pages that did not change, then compares both groups over the same window. If the treated group moves and the controls don't, the change is a plausible cause.

Lab studies are a third type. They place documents directly in a model's context and test edits. They can show cause, but only inside that setup, which is why lab GEO results often fail to transfer to live engines.

The July 2026 critical survey of 45 GEO studies put the gap 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." Most headline AI SEO research sits on the observational side of that line.

A confounder is a third factor that drives both the trait and the outcome. In AI search, it is usually brand size.

Large brands rank higher, earn more links, collect more reviews and refresh pages more often. They also get cited more. Any trait that rides with size will correlate with citations, whether or not it does anything.

The concentration is real. A 2025 arXiv study of AI search citing patterns found the top 20 news sources accounted for 67.3% of all citations for OpenAI models (data from March to May 2025). Semrush's list of 36 brands visible across all four surfaces it tracked includes YouTube, Google, Reddit and Amazon. That finding is brand size.

Most studies cannot control for it. The usual controls, Ahrefs DR and Moz DA, are vendor scores built from backlinks and traffic. They are proxies for the confounder itself, so conditioning on them does not remove it.

Seer's freshness study shows the pattern. It found 75% of cited pages were updated within a year, but it had no group of non-cited pages. If most of the web carries a recent date, the number says nothing.

The next move: only trust comparisons between like-sized brands.

Ahrefs found that 37.9% of URLs cited in AI Overviews also appeared in the first 10 results (March 2, 2026). The organic-only figure was 37.1%. Roughly 31% came from positions 11 to 100 and roughly 31% from beyond 100.

An earlier Ahrefs study put the top-10 share near 76%. It is tempting to call that a collapse. It isn't a measured trend: Ahrefs changed its parsing method between the two studies, so the endpoints come from different instruments. The 38% stands. The decline does not.

Two more limits apply:

• URL-level matching. A large site cited through a URL that doesn't rank counts as a miss, even if its other pages rank.
• An unresolved conflict. BrightEdge (September 2025) reported 16.7% of citations from the top 10, with no published sample size and an undefined "overlap". Don't cite the two as agreeing.

What it means: organic rank is associated with AI Overview citations. It is an observational proxy, not proof that ranking causes citation. For the wider picture, see what evidence supports AI Overviews ranking signals.

Ahrefs also ran one of the few matched-control studies in this field. It tracked 1,885 pages that added schema, each paired with three matched controls. Its conclusion: "Adding schema produced no major uplift in citations on any platform." AI Overview citations fell a statistically significant 4.6%, while AI Mode rose 2.4% and ChatGPT 2.2%.

That is strong design. It also has a disclosed boundary. Every treated page already had 100+ AI Overview citations before any schema was added. So the result covers pages already in the consideration set. Whether schema helps a page enter that set is untested, not refuted.

Don't overcorrect. Google says existing markup does no harm, so ripping it out isn't supported either. The honest reading: schema is not a lever for AI answers, based on the best test available. This is also a lesson in why GEO studies disagree: samples differ, and the exclusions decide what a result can cover.

Otterly's test is useful for what it observed: Google removed both sites while Copilot kept citing them. With one site per arm, it cannot isolate the author box from everything else that differed.

Run every new number through these questions before it shapes a budget:

1. What was compared to what? No comparison group means the number describes, it doesn't explain.
2. Was brand size controlled? DR or DA as a control doesn't count.
3. Did the method change between data points? A trend needs the same instrument at both ends.
4. Who was excluded? A result on heavily cited pages may not apply to a brand at zero.
5. Are the key terms defined? Seer's review-profile study never defined "citation rate". Treat undefined metrics as unusable.
6. Was a live engine queried? Or were documents placed straight into the context window?
7. Which engine, which date, retrieval on or off? Results from one surface rarely transfer to another.
8. Who ran it, and what do they sell? Vendors, including us, have a motive. That isn't disqualifying. It is why replication matters.

A number that fails three or more of these is a talking point, not evidence. Our AI SEO statistics you can actually trust applies this filter to the figures that circulate most.

• Deciding whether to spend on a change: use matched-control results, then check whether your pages resemble the sample.
• Describing the landscape: observational stats are fine, as long as you say "associated with", not "causes".
• Understanding how models pick among retrieved pages: lab studies help, but don't expect live gains from them.
• Setting access rules: use first-party documentation. OpenAI states its crawler settings are independent of each other, and Google says AI Overview eligibility requires indexing and snippet eligibility. These are the parts you control.
• Knowing what AI says about you: measure it directly, engine by engine. The free checker on this site gives a snapshot for up to three buyer questions.

For platform claims, start with what vendors disclose. Our guide to what OpenAI discloses about ChatGPT ranking factors is a short read for that reason.

Are AI SEO statistics from vendors trustworthy? Some are, if the method is published. Judge the design, not the brand. A vendor matched-control study with a disclosed sample, like Ahrefs' schema test, says more than a large observational dataset with undefined terms. Check what was compared, who was excluded, and whether the vendor sells the tactic the number supports. A motive doesn't make data wrong, but it raises the bar for replication.

Does ranking in Google's top 10 get me into AI Overviews? Not reliably. Ahrefs found 37.9% of AI Overview citations came from top-10 URLs in March 2026, with the rest split across lower positions. That is an association, not a causal path. Google's documented requirement is that a page is indexed and eligible to show with a snippet. Meeting that makes a page eligible, and Google says serving still isn't guaranteed.

What does research say about ChatGPT ranking factors? Very little that is causal. As of October 2026, OpenAI's public disclosure is that ChatGPT ranks search results using "multiple factors" and that placement is not guaranteed. Nothing is enumerated. What OpenAI does document is access: OAI-SearchBot controls ChatGPT search eligibility, and GPTBot controls training use only. Lists of ChatGPT ranking factors beyond that are usually correlational guesses.

Why can't a big dataset fix the correlation problem? Size reduces random noise, not bias. If cited pages are mostly from big brands, a million examples still show big-brand traits. Without a comparison group of similar non-cited pages, or a design that holds brand size constant, more data only makes the same confounded pattern look more precise. Matched controls or like-sized comparisons address it. Sample size alone does not.

Is a correlational study ever useful? Yes, for description. It tells you where citations land today, which surfaces concentrate on a few sources, and how engines differ. It is a good way to form a question worth testing. Where it fails is as a playbook. Treat a correlational finding as a hypothesis, then look for a matched-control test before you spend money acting on it.