Should I trust GEO case studies that show big citation gains?

Most GEO case studies don't show what their headline says. The reason is structural, not a matter of honesty.

A critical survey of 45 GEO studies (arXiv:2607.14035, July 2026) concluded that "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." In plain terms: research can show that editing a page changes how a model uses it once the page is already in front of the model. Nobody has shown that a GEO technique gets a page in front of the model in the first place.

That gap is where most case studies break. Run any claim through four questions:

1. Was there a control group? Matched pages that didn't get the change.
2. Was brand size accounted for? Big brands get cited more regardless of tactics.
3. Was it a live engine or a lab? Lab engines often hand the model the content directly.
4. How is the metric defined? If the number has no stated denominator, you can't read it.

If a study fails any one of these, treat its gain as a hypothesis. Our red flags guide covers the specific claims that fail most often.

The most quoted number in GEO is "up to 40% more visibility." It comes from the original GEO paper (Aggarwal et al., arXiv:2311.09735, KDD 2024).

The authors built their own two-step engine and supplied the source documents in the model's context. Even the one test on Perplexity uploaded source text as file attachments, because Perplexity didn't let users specify source URLs. No part of the study touched how a live engine finds pages.

The metric also inflates the result. Position-Adjusted Word Count counts words from a source that appear in the answer, and the interventions added words to the source. The paper reports a 30 to 40% band across three methods, not the itemised per-method figures often quoted. We break this down in the 40% GEO claim, examined.

Later, better-designed studies found the opposite. C-SEO Bench (arXiv:2506.11097, NeurIPS 2025) found most conversational SEO methods "largely ineffective" and often harmful to a document's ranking. The "SEO" arm that did work simply moved the target document to position 1 in the model's context, which no publisher can do. See our C-SEO Bench summary for the method.

A before-and-after chart with no control tells you almost nothing. Engines change models, retrieval behaviour and citation formats often. A gain over three months could be the engine, not the tactic.

A proper test compares treated pages against matched pages that weren't changed. Ahrefs ran one: 1,885 pages that added schema, each matched to 3 control pages. It found no major uplift on any platform, and a statistically significant 4.6% fall in AI Overviews citations. Ahrefs also disclosed its boundary: every treated page already had 100+ AI Overview citations. That honesty is what good studies look like.

Brand size is the biggest confounder in this field. A study of AI search citing patterns (arXiv:2507.05301) found the top 20 news sources accounted for 67.3% of citations for OpenAI models. If a case study features a well-known brand, its citations may reflect the brand, not the work.

Watch for jumps that sit exactly at a presence-versus-absence line, such as "brands with a review profile were cited far more." Brands with no profile tend to be small. The gap may be measuring company size.

A percentage is only as good as its denominator. Before trusting a number, find out:

• What counts as a citation. A linked source, a brand mention, or a quoted span? Engines cite differently, so totals across platforms aren't comparable.
• Which prompts were asked. Trackers count answers to prompts the vendor chose. The Otterly 2,000 AI blogs experiment logged 3,934 citations for one site, while Microsoft's own report credited it with 42,817. Time windows may differ, but any prompt set is a sample.
• Whether search was on. ChatGPT and Claude decide per query whether to search. A score that mixes searched and unsearched answers averages two different systems.
• Whether cited means used. OpenAI documents that the sources a model consulted usually outnumber the ones it cites.

If the study won't publish the prompt list or define the metric, you can't check the result.

Use this on any case study, vendor deck or LinkedIn post:

1. Is there a matched control group, and how many pages are in each arm?
2. Were treated and control sites similar in brand size?
3. Did the test run on a live engine, with real retrieval, or on content placed in context?
4. Is the metric defined, with a denominator?
5. Are the prompts, dates and model versions published?
6. Was search on, off, or mixed?
7. Who ran it, and what do they sell?
8. Has anyone independent replicated it?

A study can pass every item and still describe only one engine at one date. Results from 2025 models haven't been retested on the current generation as of October 2026.

Ask vendors for evidence on the levers that do hold up. These are access levers: content in the HTML before JavaScript runs, per-bot robots.txt rules, a firewall or CDN that lets AI crawlers through, and snippet controls set on purpose. They work by removing blocks, and you can check each one yourself.

Be wary of anyone promising a citation lift. Our page on agencies that guarantee AI citations explains why no one can promise that. For a wider view of what studies support, see what AI SEO research actually shows.

If you want a baseline before talking to any vendor, the free checker on this site shows whether AI engines name you for up to three buyer questions. It's a snapshot, not proof of cause.