US market research

Did AI-Cited Pages Use More GEO? Comparing 265 Cited and 206 Uncited Pages

Published

This article was first published in Korean by Intrix Lab on October 9, 2026.

We did not establish a difference in these data. As of 2026-10-07, we fetched 265 pages cited in AI answers and 206 pages that AI had read through search but did not cite, then counted 10 GEO-style surface patterns. Average pattern scores were 2.90 for cited pages and 2.69 for uncited pages. The 95% interval for the difference ranged from -0.08 to +0.48, including 0. This comparison produced no evidence that pages with added FAQs or summary blocks were cited more often.

What concern did the paper raise?

The GEO-Flag paper, posted on arXiv in August 2026, argued that pages modified for generative search could receive disproportionate exposure relative to their authority or relevance, amplifying misinformation. The researchers built a benchmark of 3,200 items using 8 GEO methods and evaluated detectors. Applying a detector to 10,095 pages from actual search results, they reported that 8.90% were classified as GEO and that the share rose from 7.02% in 2024 to 16.36% in 2026.

Do pages AI actually selects contain more signs of GEO? Intrix checked by comparing cited pages with uncited candidate pages within the same answers.

How did we compare them?

From 648 Korean-language answers generated by one model, we selected 300 cited pages and 300 uncited candidate pages, evenly distributed across industries. The industries were GEO agencies, dermatology, plastic surgery, legal services, web development, marketing, and video production. Uncited candidates were pages AI had read through search but had not cited in any answer.

Because the paper's detector was not publicly available, we scored 10 GEO-style surface patterns, such as FAQs, summary blocks, quotations, and source attribution phrases, using Intrix's criteria. We fixed the criteria and primary statistical test before fetching the pages. We successfully retrieved 265 cited pages and 206 uncited pages.

Did cited pages contain more patterns?

PatternCited pagesUncited pagesOdds ratio (95% interval)
FAQ46%40%1.28 (0.89–1.86)
Quotations37%29%1.40 (0.95–2.08)
Source attribution phrases19%14%1.48 (0.89–2.45)
Summary blocks7.5%4.4%1.79 (0.80–4.01)
High numerical density31%37%0.75 (0.51–1.11)
Pattern score of 3 or higher55%54%1.03 (0.71–1.48)

Quotations, source attribution phrases, summary blocks, and FAQs leaned slightly toward cited pages. Pages densely packed with numbers were more common among uncited pages. However, every interval crossed 1, making the differences indistinguishable from chance. Recalculating after matching website types, such as business, public-sector, and news websites, yielded a similar score difference of +0.17.

Does that mean GEO patterns are useless?

These results do not support that conclusion. This comparison addressed only whether AI selects pages with more patterns for citation from among pages it has already read. We did not measure whether patterns affect a page's chances of appearing in search results. The uncited candidates were also pages AI had already read, so both groups had passed the search stage.

There is one practical takeaway for businesses. Formatting work alone, such as adding FAQs or summary blocks, is not a reliable basis for expecting citations. The first priority is whether the page contains content that answers the question. To check which pages AI reads in your industry and which it selects for citation, you can start with a first measurement.

How far can these results be trusted?

Our rules differ from the paper's detector. They do not catch stealth GEO that leaves no surface traces or content edited manually by people. The answers came from just one model in Korean, and the industry mix included questions about GEO agencies.

The page retrieval failure rate was higher for uncited pages, at 31%, than for cited pages, at 12%, so the remaining pages may be biased. We fetched the pages in October rather than August, when the answers were generated, so their content may have changed in the meantime. Given the sample size, we could detect only differences corresponding to odds ratios of roughly 1.6 or higher. This analysis does not guarantee any particular citation or ranking.

Frequently asked questions (FAQ)

Q. Does adding an FAQ increase AI citations?
That was not established in these data. FAQs appeared on 46% of cited pages and 40% of uncited pages, but the difference was indistinguishable from chance. First, check whether the page contains an answer that matches the question.

Q. Doesn't including more numbers lead to more citations?
This comparison pointed in the opposite direction. Pages with high numerical density accounted for 31% of cited pages and 37% of uncited pages. However, this difference was also not statistically significant, so we do not use it as a reason to remove numbers.

Q. Are pages modified for GEO actually becoming more common?
The paper reported that the share of pages classified as GEO in actual search results rose from 7.02% in 2024 to 12.80% in 2025 and 16.36% in 2026. We did not independently verify this trend with our data.

← Back to insights