← Back to GEO Academy
Trends

A GEO Study Flagged 8.90% of Pages: What Does That Really Measure?

A careful reading of the August 2026 GEO-Flag study, including what its 8.90% and 16.36% estimates mean, their limits, and how teams can build a defensible AI search visibility baseline.

Published 08/27/2026 8 min read
GEO researchGEO prevalenceAI search visibilityGEO monitoring

A GEO Study Flagged 8.90% of Pages: What Does That Really Measure?

GEO-optimized content is no longer merely hypothetical. Measuring how common it is, however, requires more care than repeating one percentage.

The August 20, 2026 revision of *GEO-Flag: Detecting and Measuring GEO-Optimized Web Content* classified 898 of 10,095 usable pages from released Google Search and Gemini-grounded retrieval results as showing GEO signals. That produced an estimated prevalence of 8.90%. Among pages with a parseable declared modification date in 2026, the estimate was 16.36%.

Neither figure means that 8.90% of the web has deliberately adopted GEO.

How the 8.90% estimate was produced

The researchers sampled 1,000 informational queries from the ORCAS real-user query collection. They then used previously released URL results from conventional Google Search and Gemini-grounded retrieval. Of 13,985 distinct URLs, 10,095 yielded usable content after normal fetching, browser rendering, PDF extraction, and permitted recovery steps.

The study applied a detector trained on a controlled GEO benchmark and reported:

  • 898 of 10,095 unique pages were flagged, or 8.90%, with a 95% Wilson interval of 8.36% to 9.47%;
  • page-level estimates were 8.14% for Google Search and 9.09% for Gemini;
  • across 965 queries with usable pages from both channels, the query-balanced Gemini–Google difference was 1.43 percentage points, with a paired-bootstrap interval of 0.46 to 2.41 points;
  • only 19.57% of usable pages had a parseable `dateModified` value; within that subset, pages declaring a 2026 modification date had a 16.36% flag rate.

These are estimates for a defined sample and pipeline, not official platform statistics.

Why detection rate is not adoption rate

Live webpages do not carry a gold-standard label saying whether a publisher intentionally performed GEO. A detector can miss subtle interventions or misclassify unrelated edits. The paper accordingly describes detected signals, not proven publisher intent.

A modification date is also not a GEO implementation date. It is publisher-supplied metadata, and it was available for only about one-fifth of analyzed pages. The increase seen from 2024 to 2026 is a descriptive association in the sample, not proof that GEO adoption grew at that rate.

The retrieval channel changes the denominator as well. Google and Gemini surface different source pools, page types, and numbers of URLs. Their 8.14% and 9.09% estimates should not be simplified into a claim that one platform is inherently easier to manipulate.

Build a brand-specific baseline instead

The practical lesson is methodological. A useful AI visibility baseline should preserve its measurement context:

  1. Freeze the question set and record language, market, platform, interface, and timestamp.
  2. Save the full answer, citation URLs, brand position, and relevant page version—not just screenshots.
  3. Calculate results by platform and intent before considering any combined view.
  4. Report sample size, unavailable pages, and failed observations so the denominator remains visible.
  5. Describe changes as results within the monitoring window until repeated measurements support a trend.

GEO Radar helps teams evaluate brand visibility, recommendation differences, and competitors across AI platforms using structured monitoring. At [https://www.georadar.top](https://www.georadar.top), the sound use of these capabilities is to establish your own repeatable baseline, not to apply a paper’s prevalence estimate to an individual brand.

The useful conclusion

GEO-Flag provides early empirical evidence that GEO-like interventions are visible in a meaningful share of the study’s search-result pages. Just as importantly, it models good restraint: detection is not intent, declared modification time is not implementation time, and a sample estimate is not a web-wide adoption rate.

For brand teams, a defensible GEO program begins with a stable denominator, repeated observations, and versioned evidence.

Sources for this article