Passing this check means your site is eligible to be cited by AI tools. It does not mean they will recommend you.
GEO (generative engine optimization) starts with your website, not with new content. First confirm the site can be crawled, understood and trusted. Below are 20 general checks, extra items for overseas and China, and a note on which advice has evidence behind it.
🔎 1. Why check the site first
Google's documentation says there are no additional requirements to appear in AI Overviews or AI Mode, and no need for new machine-readable files, AI text files, markup or special schema.org data. A page needs to be indexed and eligible to show with a snippet. (Google Search Central: AI features and your website)
So the foundation is still the classic basics, plus facts that are clear, verifiable and consistent.
Overseas sites also need correct language annotations, privacy and cookie compliance, and a company profile that matches across third-party platforms.
🧪 2. Which advice has evidence
| Grade | Meaning | Example |
|---|---|---|
| A | Stated by a protocol, standard, official documentation or law | robots.txt rules, sitemap specification |
| B | Public search-engine guidance or industry consensus; effect on AI answers not quantified | Clear navigation, consistent facts, complete content |
| C | Experience or hypothesis with no public evidence yet | Particular phrasing patterns, Q&A formats |
| X | Data currently shows no effect | llms.txt, see section 6 |
✅ 3. Twenty general checks
🚪 Can it be accessed?
- Important pages open normally, with no large numbers of dead links.
- robots.txt does not block core pages. A missing file is not an error; a server error should be fixed.
- A sitemap exists and lists only pages that load normally.
- No page carries an unintended noindex.
- Your firewall or CDN does not return 403 to search and AI crawlers across the board.
- The certificate is valid, the site runs on HTTPS, and http redirects once to https.
🧠 Can it be understood?
- Core text is in the raw HTML, not produced only by scripts.
- Each page has a unique title and a clear main heading; descriptions are full sentences, not keyword lists.
- Navigation lets people and crawlers find every core offering, and menu names are self-explanatory.
- Service pages say what it is, who it is for, how it works, what it costs and what the limits are.
- The first screen contains one sentence that can be quoted as a definition.
🤝 Can it be trusted?
- Company name and contact details are clear in the footer and on an About page, and written the same way everywhere.
- Key numbers state source, date and basis; case studies state their scope and do not exaggerate.
- A privacy policy and terms exist and are linked next to every form that collects information.
- Articles show an author and publish and update dates.
- No unsupported absolute claims; promises state their conditions, exceptions and remedies.
🔁 Can it be verified?
- The same facts (name, price, scope of service) match across pages and languages.
- Independent official profiles back you up, such as company registries, official accounts and industry directories.
- You test regularly with a fixed set of questions in the models you care about, and record whether your brand is mentioned and described correctly.
- You re-test after every change instead of trusting a feeling.
🌍 4. Extra items for overseas sites
- Each language has its own URL; language versions point to each other with hreflang, and prices and terms match across versions.
- Privacy policy and terms of service. For EEA and UK visitors, cookie consent that can be declined, no non-essential scripts before consent, and full page content after declining.
- Matching company information on LinkedIn, Crunchbase, Product Hunt and similar platforms.
- If your brand name is generic, use a fixed "brand + category" form so it is not drowned out by products with the same name.
- Use IndexNow so Bing-based search learns about updates faster.
🇨🇳 5. If you also sell in China
- Show your ICP filing number in the footer and link it to the official lookup page; the filing entity should match the site's operator (check whether this applies to you).
- Avoid the extreme wording that Chinese advertising law restricts, and cite data with source and validity period.
- Check access speed from several mainland networks, and avoid third-party resources that are unreachable there.
⚠️ 6. Three common myths
- "An llms.txt file gets me cited." Ahrefs' analysis of 137,210 domains (published June 2026) found about 38,000 with a valid llms.txt, and 97% of those files received no requests in May 2026. AI retrieval bots made up about 1.1% of requests. Public data does not support the claim. (Ahrefs: llms.txt study of 137K sites)
- "FAQ markup improves recommendations." The value is in having clear questions and answers on the page, not in the markup.
- "More keywords is better." Modern retrieval looks at meaning and facts; stuffing only hurts readability.
🧰 7. How to use this list
Go through it yourself, then hand uncertain items to a tool or a specialist. A machine can tell whether something exists. A person has to confirm that qualifications, case studies and legal compliance are genuine.
Passing means you are eligible to be cited, not that you will be recommended.
This article reflects public sources and our own checking practice. It makes no promise about any search engine's or AI model's recommendations. For legal and compliance matters, rely on current regulations and professional advice.
Originally published on ShareHub: GEO Website Self-Check: 20 Checks Before You Invest in GEO Content (2026-10-18). ShareHub is an early-stage traffic matchmaking platform that connects brands with micro-creators.
Top comments (3)
Peter, the evidence grading table is the most original part of this article. Most GEO guides put everything on the same level.
One data point for the "access" group. I tested 636 sites at the HTTP level to see which AI crawlers actually get blocked, not just what robots.txt says. About 20% block at least one AI crawler: GPTBot 17.7%, PerplexityBot 15.2%, Applebot 13.8%. 73% of generic blocks (all bots, including browser) are on Cloudflare. So check #5 is more common than it looks if you only check robots.txt, and your two-site sample probably doesn't catch it.
On the "verified" group CiteScope raised: across 341 pages, sites that block at least one AI bot in robots.txt have a citation rate of 28.4% vs 47.0% for those that don't. An 18.6 percentage point gap. That quantifies what you say in the article: eligible doesn't mean cited.
The data comes from an open-source GEO audit tool I'm building (geo-optimizer-skill on GitHub), so the dataset is verifiable.
把建議分成 A / B / C / X 證據等級這一點很有用,尤其 llms.txt 那段可以直接幫人省下不少無效工作。實務上我覺得最容易被忽略的是最後那組「可驗證」:很多網站頁面內的名稱、價格、服務範圍寫得不一致,換個語言版本又差更多,AI 自然不敢引用。還有一點是固定題目重測,改版後如果不重新跑一次,很容易以為有改善其實沒有。想請教你目前看到的案例裡,20 項之中哪一項最常不合格?是爬取、理解,還是信任那一組?
Thanks for reading. Honest answer: I have only run these checks on our own two sites, so this is a small sample, not a survey. The items that failed most were in the "understood" group (duplicate page titles, thin pages) and the "trusted" group (missing footer and policy details). Crawl access was mostly fine. I agree the "verified" group is easy to miss, and re-testing after each change is the part people skip most. Which of the 20 would you check first on your own site?