The short version
- In an audit of 1,357 Gemini citations for 156 Tokyo hotel questions, researchers at Blossom AI compared 14 hotel websites (Zhu and Chang, March 2026).
- Hotels that Gemini cited directly scored 8.6 out of 15 on content depth; hotels it skipped scored 3.4, and every hotel scoring 6 or more was cited.
- Kadoya Hotel, with no schema markup and no technical SEO, scored 13 out of 15 thanks to a 33-question FAQ and a 13-attraction sightseeing guide, and was cited.
- Hotel K5, with a 9.6 out of 10 guest rating and full technical setup, scored 5 and was never cited directly; Gemini described it using booking sites instead.
What did the Gemini hotel audit find?
Cited hotels had much deeper question-answering content than skipped ones, regardless of technical polish.
Zhu and Chang (opens in a new tab) are two researchers at Blossom AI, a San Francisco company. In March 2026 they put 156 hotel questions to Gemini 2.5 Flash with Google Search grounding switched on. Grounding means Gemini runs Google searches and cites the pages it uses. The questions covered nine Tokyo districts in English and Japanese and produced 1,357 citations.
They then audited the websites of seven hotels Gemini cited directly and seven it did not. Each site was scored from 0 to 3 on five kinds of content: an FAQ, an area guide, a blog, access information and distinctive content. The key test was depth, meaning whether each feature was rich enough to rank in Google for real traveler questions.
| Group (14 hotels) | Average depth score |
|---|---|
| Cited directly by Gemini | 8.6 out of 15 |
| Not cited | 3.4 out of 15 |
The split was clean. Every hotel scoring 6 or above was cited, and every hotel below 6 was not.
Why did the hotel without schema beat the one with it?
Because Kadoya’s pages answered real questions in depth, while K5’s features were present but thin.
Hotel K5 is a design boutique hotel with a 9.6 out of 10 rating on Booking.com. It has bilingual content, Schema.org markup, a neighborhood page and an FAQ. Each of those features is brief, so it scored only 5 out of 15. Gemini named K5 in three answers but drew its facts from Expedia, Hotels.com and an editorial site, not from K5’s own website.
Kadoya Hotel is an independent with no schema markup and no technical SEO. It scored 13 out of 15. Its 33-question FAQ, 13-attraction sightseeing guide and regular blog create more than 20 indexed pages that answer traveler questions directly.
The authors draw a two-step lesson. First, a hotel’s site needs content deep enough to rank in Google for relevant questions. Second, that content has to answer the question better than competing booking or editorial pages. The authors conclude that brand prestige and technical sophistication did not determine direct citation; content depth did.
How often does Gemini cite a business’s own website?
Not often. Booking sites and other intermediaries took most citations, even when hotels were named.
Online travel agencies such as Booking.com and Expedia took 55.3% of all 1,357 citations. Hotels’ own websites took 8.2% of English citations and 11.0% of Japanese ones. Japanese hotel sites tend to carry deeper neighborhood, transit and area content, which the authors link to their higher direct citation.
How the question was asked mattered too. Questions about atmosphere or experience drew 55.9% of citations from sources other than booking sites, against 30.8% for plain booking-style questions. Hotels’ own sites held a steady share in both, which suggests deep pages help across many kinds of questions.
Gemini is more open to brand-owned pages than some other assistants. A comparison of AI engines by Chen and colleagues (opens in a new tab) at the University of Toronto found Gemini drew 21.2% of its sources for niche brands from brand-owned sites. ChatGPT drew just 4.9%.
Does technical SEO matter at all for AI citations?
Ranking still matters; schema on its own shows little sign of mattering.
The hotel authors are clear that Kadoya’s pages still had to rank in Google to be found. Gemini’s grounding runs on Google Search, so a page that never appears in search never reaches the second step. Kadoya’s 20-plus indexed pages, each answering traveler questions directly, cleared that first step without any technical SEO.
Our own data supports both halves. In our study of 3,096 pages ranking in Google’s top 10, AI Overviews cited 41.7% of pages in positions 1 to 3. That is about twice the rate at 7 to 10. Once pages on the same search were compared, no schema type held up as a reliable predictor of citation.
Rankings for the question as typed are only part of the story. In our four-assistant study, 16.6% of Gemini’s citations ranked in Google’s top 10 for the buyer’s question. And 49.8% of Gemini’s citations sat outside Google’s top 100 for the question and for the related searches the study tracked. Pages that cover many specific questions give an assistant more ways to find them.
Is depth the same as length or an FAQ page?
No. Depth means covering real questions with specific answers; format and word count alone do little.
K5 had an FAQ and a neighborhood page and still failed, because both were brief. A large study of ChatGPT, Google and Perplexity citations by Zhang and colleagues (opens in a new tab), three independent researchers, found the same. Pages in a question-and-answer format had 5.74% less influence on answers, on the authors’ measure, than other pages. The pages that shaped answers most held definitions, figures, comparisons and how-to steps.
A 2026 controlled test by Vishwakarma and colleagues (opens in a new tab) at Sprinklr, a software vendor, points the same way. Across six AI models, pages with fuller specifications and more comprehensive analysis were cited first more often, while formatting-only changes had no consistent effect.
What should you do about it?
Invest first in pages that answer your buyers’ specific questions in depth, and treat technical markup as a supporting task.
- List the concrete questions buyers ask before they choose you, such as access, timing, prices, comparisons and local details.
- Answer each in depth on your own site, with specifics no booking or review site has.
- Build depth, not features. A long, organized FAQ beats a short one; one rich guide beats several thin pages.
- Publish in the languages your buyers search in. Japanese searches, where hotel sites carry deeper content, produced more direct hotel citations than English ones.
- Keep basic technical health so pages can be crawled and ranked, but do not expect schema alone to earn citations.
- Check which sources Gemini cites when it mentions you. If it names you but cites intermediaries, your own pages are not answering the question well enough.
If you want help running that check across assistants, see our generative engine optimization service.
What does the research not tell us yet?
The hotel finding is exploratory, and its authors say plainly that it shows correlation, not causation.
- The depth audit covered only 14 hotels in one city, scored by the authors on their own scale.
- Better-known hotels may both invest in content and be cited for brand reasons; the study cannot separate the two.
- It tested one engine, Gemini 2.5 Flash, in March 2026. ChatGPT, Perplexity and Claude use different search systems.
- Each question was asked once; a repeat of 20 questions showed citations vary from run to run.
- The authors work for a company and propose, but did not run, a test that adds content to a hotel site and tracks citations.
Frequently asked questions
Does Gemini need schema markup to cite a website?
No. In the Tokyo hotel audit, Kadoya Hotel had no schema markup and was cited directly, while a hotel with full markup but thin content was not.
What kind of content does Gemini cite from business websites?
Deep pages that answer specific customer questions. Cited hotel sites scored 8.6 out of 15 on question-answering depth, against 3.4 for hotels Gemini skipped.
Can a small business compete with big brands in AI search?
Sometimes, on specific questions. An independent hotel with a 33-question FAQ was cited directly, while a highly rated boutique rival reached Gemini answers only through booking and editorial sites.
Is an FAQ page enough to get cited by AI?
No. Hotel K5 had an FAQ and was not cited because it was brief, and a separate study found question-and-answer formatting alone went with 5.74% less influence on answers.
Sources
- Zhu and Chang (2026), The End of Rented Discovery: How AI Search Redistributes Power Between Hotels and Intermediaries (opens in a new tab), arXiv:2603.20062.
- Chen, Wang, Chen and Koudas (2025), Generative Engine Optimization: How to Dominate AI Search (opens in a new tab), arXiv:2509.08919.
- Zhang, He and Yao (2026), From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms (opens in a new tab), arXiv:2604.25707.
- Vishwakarma, Kumar and Jamidar (2026), What Gets Cited: Competitive GEO in AI Answer Engines (opens in a new tab), arXiv:2605.25517.
- Underneath (2026), What pages cited by AI Overviews have in common: 3,096 pages
- Underneath (2026), Do ChatGPT, Gemini, Perplexity and Claude cite pages that rank?