Does GEO work? We read the studies behind the slides.

The paper behind the famous 40%

GEO: Generative Engine Optimization by Aggarwal and colleagues was published at KDD 2024. The authors built a benchmark of 10,000 queries. For most of the experiments, they fetched the top five Google results for each query and had GPT-3.5 write an answer from them. Then they rewrote one of the five sources in different ways and measured how much of the answer was attributed to it.

Their best methods, citing sources, adding quotations and adding statistics, achieved “a relative improvement of 30-40%” on a metric called Position-Adjusted Word Count. Keyword stuffing offered “little to no improvement”. In a smaller test of 200 queries on Perplexity, quotation addition performed best with a 22% improvement on the same metric.

One more finding gets less airtime: the gains were uneven. Citing sources helped the fifth-ranked page a lot, while the top-ranked page’s visibility fell by about 30%. In a fixed set of five sources, visibility is a shared pie. If everyone adds statistics, nobody stands out because of the statistics.

What later research adds

A July 2026 critical survey of GEO research reviewed the literature from 2023 to 2026. It concludes that “topical relevance and context position are the most reproducible levers”, that generic heuristics transfer poorly, and that “no reviewed technique shows a stable, longitudinal, cross-platform causal effect” on organic discoverability. It is a preprint, so it has not necessarily been peer reviewed, but its conclusion matches what the individual studies below show.

In plain words: being the best answer to the actual question still matters most. Formatting tricks can help a page that is already in the running. They rarely get a page into the running.

Schema markup: the evidence is thinner than the advice

Sources: Ahrefs, GEO-16, searchVIU. The only controlled comparison here, Ahrefs, found no meaningful lift. It only looked at pages that were already being cited, so it says less about pages starting from zero. searchVIU itself notes that markup may still matter earlier in the pipeline, for indexing and search.

Google’s position is simple: there is “no special schema.org structured data” needed for AI Overviews or AI Mode. Our take: keep accurate structured data for the search features it already supports. Do not expect it to win a recommendation, and put anything important in visible text.

The ground keeps moving

Profound compared the domains cited for about 80,000 prompts per platform, one month apart, in June and July 2025. It found that “roughly 40-60% of the domains cited in AI responses will be completely different just one month later” (Profound). That is a vendor dataset and a single comparison, but it is a large one.

Answers to the same prompt vary too. In a SparkToro and Gumshoe study, 600 volunteers ran 12 prompts through ChatGPT, Claude and Google’s AI 2,961 times. The authors estimate a less than 1 in 100 chance of getting the same list of brands twice. Their conclusion: a visibility percentage across many prompts and runs is a reasonable metric; a single “ranking position in AI” is not.

This is why a before-and-after test on one prompt, run once, proves almost nothing. Our guide to reading AI visibility changes explains how to repeat a test properly.

Does ranking in Google still matter?

For Google’s own AI features, yes, but less tightly than it used to. Ahrefs found in July 2025 that 76% of AI Overview citations came from the top 10. Its March 2026 update found 38%, linking the change to query fan-out, where Google searches related sub-questions. The method changed between the two studies, so the comparison is indicative, not like for like.

Other assistants overlap even less with Google. An August 2025 Ahrefs analysis found that only 12% of links cited by ChatGPT, Gemini and Copilot appeared in Google’s top 10 for the same query. So SEO helps, but a strong Google ranking does not guarantee a ChatGPT mention.

What we would actually do with this

• Fix relevance first. Find the buying questions where you should appear and do not, and check whether you have a page that genuinely answers them.
• Add evidence where it belongs: real figures, named sources and quotes from people who know the subject. That makes a page more useful, and it matches what the original paper found.
• Keep the technical basics sound, but do not expect schema or special files to do the heavy lifting.
• Measure with repeated runs across many prompts, and compare like with like. A single screenshot is an anecdote.

If that sounds a lot like good content work with a new measurement layer, that is roughly what the evidence supports. Our SEO, AEO and GEO guide turns it into a page brief.