Are SEO and GEO Really the Same? We Tested 48 Queries

This article was first published in Korean by Intrix Lab on August 14, 2026.

87% of the sites cited by AI were outside Google's top 10 search results (measured on 2026-08-14). "Isn't AI search optimization ultimately just SEO?" This is one of the industry's most common objections. The logic is simple: if AI searches the web to answer, the sites that rank highest on Google should also receive the most AI citations. It sounds reasonable.

So we measured it. We submitted the same 48 questions to Google and two generative AIs, then compared the sites each returned (measured on 2026-08-14, 480 calls, 0 failures).

Here is the result. For one model, 87% of cited sites were entirely absent from Google's top 10. For the other, it was 63%. The two AIs also overlapped with each other less than either overlapped with Google.

How we measured it

The questions were 48 local recommendation queries, such as "Recommend good dermatology clinics in [location]" (3 sentence patterns × 16 locations). We chose wording that people might actually type into a search engine or AI.

With search enabled, we asked each AI each question multiple times. A single query can lead you to mistake a chance result for consistent performance. In our earlier measurements, 40% of the clinics cited by AI changed in just four days.

The comparison unit was the domain. Counting pages would treat different pages on the same site as separate results, understating the overlap.

Does AI use the sites Google ranks highly?

Very little. But the degree varies substantially by model.

For Model A, 87.1% of cited domains were outside Google's top 10. Even when we expanded the range to the top 30, 80.0% remained outside it. The overlap index (Jaccard) was 0.087.

For Model B, 62.5% were outside the top 10, with an overlap index of 0.294. It was 3.4 times closer to Google than Model A.

For comparison, research outside Korea has reported that 53% of Google AI Overviews citations fall outside the organic top 10. Our measurements were more extreme. This appears to suggest greater separation in the context of Korean-language queries, specifically local recommendations.

The gap widens when platforms such as directories and portals are excluded. For Model A, 90.0% were outside Google's top results, compared with 72.1% for Model B. Platforms had been increasing the overlap because they also tend to rank highly on Google.

Are the two AIs similar to each other?

No. The overlap index for the two models' cited domains was 0.080, lower than either model's overlap with Google. It was also below the cross-engine figures reported outside Korea (0.11–0.18).

In other words, optimizing for one AI does not carry over to another. Describing "AI search optimization" as a single task is itself inaccurate.

Why does this difference occur?

The reason becomes apparent when the most-cited site lists are placed side by side.

Google's top search results repeatedly feature clinic booking and consultation platforms. The sites cited by Model B are almost the same platforms. In contrast, Model A most frequently cited dermatology societies and professional organizations.

The suspected mechanism is simple. Model B uses Google Search directly, so it cites what Google surfaces. Model A follows a different search route and reaches different kinds of sites. Different search routes lead to different citation sources.

So the statement "AI search optimization is just SEO" is about half right for Model B and mostly wrong for Model A. Neither is the whole picture.

One in five answers has no sources at all

Something unexpected emerged during measurement. Out of 240 responses, Model B answered 51 times (21%) without searching at all. Naturally, those responses had 0 sources. Yet the answers were still over 1,700 characters long.

More interestingly, this varied by location. In Gyeonggi and Seoul, it skipped search 93% of the time. In 7 other locations, it never skipped search. It answers without looking things up for locations it already knows from training.

When we previously measured brand knowledge with search disabled, AI invented nonexistent information instead of saying "I don't know." The 21% found here is precisely that situation. For questions recommending medical institutions, one in five answers had no sources.

What should you do?

First, do not estimate AI visibility from search rankings alone. In most cases, AI did not use sites that ranked highly on Google. Conversely, it often used sites that did not rank highly.

Second, check each model separately. When the two models' citation sources differ this much, performing well in one provides no basis for assuming you will perform well in the other.

Third, start by checking which sources are being cited. In this industry, platforms and professional societies were cited far more often than clinic websites. Improving only your own website without understanding that structure means missing where citations actually occur.

Our free SEO and GEO audit measures whether your site is accessible to AI crawlers and reachable through search. However, no action guarantees a citation in a particular answer. What we can show is a measurement of the prerequisites.

Limitations of this measurement

We will state these plainly. All 48 questions concerned local dermatology clinic recommendations. This is one industry and one question type, in the specific context of commercial recommendations. We do not assume that informational questions or other industries would produce the same results.

This is also a measurement at one point in time. Models keep changing. We recently ran the same experiment with older and latest models, found that two of our earlier conclusions reflected model-generation differences, and withdrew them. These figures may also change in the next generation.

Frequently asked questions (FAQ)

Q. Which models did you measure?
Two commercial models that were the latest as of August 2026. We measured them with search enabled. Using older models mixes model-generation differences into the results, so we collected the data again using the latest versions.

Q. What does a Jaccard overlap index of 0.087 mean?
If two lists are identical, the value is 1. If they have no overlap, it is 0. A value of 0.087 means the overlap is less than one in ten sites. For comparison, cross-search-engine figures reported outside Korea are 0.11–0.18.

Q. Does this mean SEO is unnecessary?
No. In our earlier measurements, sites that were not indexed in search received almost no AI citations. Being reachable through search remains a prerequisite. This experiment says that rankings do not determine citations, rather than that search is irrelevant.

Q. Can we check our own company?
Yes. But AI answers vary between runs, so you should not judge from a single query. Repeated measurements are needed to see a pattern.