How much web content is already optimized for AI search?

How much of what Google and Gemini show looks optimized for AI?

About 9% of pages, in the one large audit published so far. Chu and colleagues (opens in a new tab), researchers at CISPA Helmholtz Center for Information Security, HPE and the University of Waterloo, built a detector for GEO: generative engine optimization, meaning edits that make a page more likely to be picked and cited by AI search engines. Our plain-language AI search glossary defines GEO and the related terms.

They ran it on the pages that Google Search and Gemini returned for 1,000 real user searches, fetched between 28 and 31 July 2026. Of 10,095 usable pages, the detector flagged 898, an estimated 8.90%. The rate was 8.14% for pages from regular Google Search and 9.09% for pages Gemini used to ground its answers.

The authors call these estimates, not measurements. Live pages carry no label saying they were optimized, and the detector makes mistakes. It also struggled more with light-touch and human-written optimization, so subtle work may be undercounted.

Is the share of AI-optimized content growing?

It appears to be, though the evidence is a trend, not proof. Among pages that declared when they were last modified, the flagged share rose from 7.02% in 2024 to 12.80% in 2025 and 16.36% in 2026. For 2026 pages it reached 13.52% on Google Search and 18.20% on Gemini.

There is a large caveat. Only 19.57% of pages had a readable modification date, and that date says nothing about when or whether optimization happened. The authors say the rise “should therefore be treated as a descriptive trend rather than evidence of increasing GEO adoption.”

Even so, the direction matters for planning. Pages that were recently touched are more likely to look optimized, so the baseline you compete against is rising, not static.

Where is AI-optimized content most common?

On commercial and creator platforms more than on reference sites, in this sample. 20.37% of the Amazon pages in the audit were flagged, against 4.38% of YouTube pages. None of the 613 Wikipedia pages was flagged, and several large health sites also had none.

The samples per site are small: 54 Amazon pages and 434 YouTube pages. The authors suggest commercial platforms have stronger reasons to optimize for AI shopping and discovery tools, but say their data “cannot establish the intent behind individual pages.”

Quality is a concern on the flagged pages. Of 6,663 citations inside them, 69.34% pointed to sources the authors rated low in verifiability, meaning weak editorial accountability or hard to check. That share was 74.15% on pages Gemini used and 45.88% on pages from Google Search. A low rating does not mean the claim is false. A related question is whether AI engines cite AI-written pages.

What other signs show websites adapting to AI?

Technical signals for AI agents remain a minority practice among large websites. In our study of 5,902 top websites, 11.5% served a valid llms.txt file, a proposed summary file for AI tools. Only 3.2% returned a Markdown version of a page when an AI agent asked for one (our agent-readable web study). Structured data is more common: 45.8% of homepages carried it.

Some content is shaped to steer AI answers more directly. In our study of AI-cited “best of” lists, 24.2% of numbered lists with an identifiable publisher ranked their own publisher first.

At the far end are hidden instructions aimed at AI tools. Scanning 1.2 billion web addresses, Khodayari and colleagues (opens in a new tab) found 15.3 thousand confirmed instances, including roughly 1.5 thousand meant to manipulate reputation. Many were long-lived: 65% of archived pages already carried them 12 months earlier. These are tiny numbers against the whole web, and in the authors’ tests AI tools followed them at most 8% of the time.

What happens when everyone in a category optimizes?

The advantage shrinks, but sitting out looks worse, at least in one simulated market. Chu and Hou (opens in a new tab) tested skincare recommendations from three AI assistants using fictional challenger brands and one real market leader.

When one challenger used authority-style marketing language, the leader’s survival as the recommended pick fell from 100% to 19.8%. When all nine challengers used it, the leader recovered to 93.8%. Each brand’s gain on the authors’ scale fell from 0.802 for the first mover to 0.007 once everyone optimized, close to nothing, and brands that did not optimize received zero recommendations.

Two cautions apply. The language tested included made-up clinical claims, which no brand should copy; our guide on how GEO can push false claims covers that risk. And this was a controlled test of three assistants in one product category, not a live market. Still, it describes the pressure the audit above suggests is building.

What should you do about it?

Assume competitors are already optimizing, and compete on verifiable substance rather than tricks. In practice:

1. Check your category. Ask AI assistants your buyers’ questions and note which pages and lists they cite.
2. Look at your recently updated pages first. That is where optimized competitors cluster, by the audit’s date trend.
3. Back claims with checkable sources. Flagged pages leaned heavily on low-accountability citations, so verifiable evidence is a way to stand apart.
4. Do not rely on self-ranking lists or hidden instructions. Self-ranking lists were 1.1% of all AI citations in our study, and hidden instructions rarely worked in testing.
5. Track whether your share of AI recommendations holds as competitors optimize, not just whether you appear once.

For help running that kind of category review, see our generative engine optimization service.

What does the research not tell us yet?

Nobody knows the true share of AI-optimized content, because no page carries a reliable label.

• The 8.90% figure comes from one detector, one set of 1,000 searches and one week of fetching. The detector can miss subtle, human-written optimization.
• The rise over time rests on the 19.57% of pages with a readable date and is not evidence of adoption.
• Per-site figures rest on small samples, such as 54 Amazon pages.
• The competition study is simulated, with fictional brands in one product category.
• No study yet links a page being flagged as optimized to whether it actually wins more AI citations in live search.

Frequently asked questions

Are my competitors already doing GEO?

Some probably are. An audit of 10,095 pages from Google Search and Gemini flagged 8.90% as optimized for AI, and 16.36% of pages updated in 2026.

Is Gemini more exposed to AI-optimized content than Google Search?

Slightly, in the one audit available. The detector flagged 9.09% of pages Gemini used against 8.14% of Google Search pages, and 18.20% against 13.52% for pages updated in 2026.

How many websites have an llms.txt file?

A minority of large sites do. In our check of 5,902 top websites, 11.5% served a valid llms.txt.

If everyone optimizes for AI, does it stop working?

In one simulated market, mostly yes for the gain, but not for the cost of opting out. Each brand’s gain fell to near zero once all nine challengers optimized, and brands that did not optimize got zero recommendations.

Sources

• Chu, Leng, Li, Shen, Shen and Zhang (2026), GEO-Flag: Detecting and Measuring GEO-Optimized Web Content (opens in a new tab), arXiv:2608.16824.
• Chu and Hou (2026), Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems (opens in a new tab), arXiv:2606.17443.
• Khodayari, Zhang, Acharya and Pellegrino (2026), Indirect Prompt Injection in the Wild: An Empirical Study of Prevalence, Techniques, and Objectives (opens in a new tab), arXiv:2604.27202.
• Underneath (2026), How many websites have an llms.txt file? 2026 adoption data
• Underneath (2026), Do websites serve Markdown to AI agents? 2026 data
• Underneath (2026), How many “best of” lists cited by AI rank their own brand first?