AI Readability of Changzhou AI-Device Websites: A 21-Company Pilot

Direct answer: On 31 October 2026 we ran a fixed search-and-fetch procedure on 21 companies from an official Changzhou list of AI terminal products. 7 had an analyzable official homepage, 13 produced no identifiable official site in the fixed search, and 1 blocked all crawlers in robots.txt, including five AI crawlers. Among the 7 homepages, 0 had a unique H1, 1 had JSON-LD, 0 declared hreflang, and 1 had a recognizable inquiry form.

This is a pilot, not an industry report and not a website ranking. We measured homepages only, with a single static fetch, no JavaScript execution, no form submission, and no real-user performance data. The sample comes from one specific list and does not represent Changzhou manufacturing. We publish the frame, search rules, metric definitions, an engine change, anonymized data and limitations so others can check the work.

What this pilot answers, and what it does not

It describes the discoverability, static extractability, structured data, language declarations and inquiry paths of companies on one official AI terminal product list. It does not describe Changzhou manufacturing as a whole, rank companies, or estimate how many AI citations or inquiries structured data would generate.

The question matters because "can AI read our website" is now a routine procurement and content question, yet public discussion is mostly tool reviews and experience posts without a traceable sample frame. We used a government product list because it has a publication date and an index number, the companies are manufacturers or hardware makers, and the list does not contain website URLs, so we could not pick samples by outcome.

Sample frame: 21 companies from one official list

The frame is the Changzhou municipal government notice "Announcement of the 2026 Changzhou AI Terminal Innovation Product List (Second Batch)", issued by the municipal industry bureau on 8 October 2026 (index 014109453/2026-00054). The list has 21 companies and 25 products; companies with multiple products count once, giving 21 units.

This is a census of the list, not a random sample: all 21 companies are included and were collected in list order over seven batches. There was therefore no step of choosing flattering companies. The limitation remains that the list covers only these companies, which vary widely in size, product type and age.

Source: Changzhou municipal government notice.

Method: finding the site, fetching it, and measuring

We first identified the official site with fixed queries, then checked robots.txt before fetching the homepage once. We recorded status, TTFB, HTML size, H1 count, JSON-LD types, hreflang, forms, mailto/tel/privacy links and other reproducible fields. Raw HTML stays off-site; public data keeps only anonymized sample IDs.

Site identification: two queries, "<legal name>" 官网 and "<legal name>" 官方网站 首页, top 10 results. A result was accepted only if the title or snippet contained the legal or brand name, the target homepage contained the name or a verifiable brand, and the domain was not a B2B directory, company registry, job board, social network, news outlet or government page. If neither query produced an acceptable site, we recorded "no official site identified". Exhibition or marketplace profiles were not counted as company sites.

Fetching: we requested /robots.txt first and checked whether our user agent was allowed on the homepage. If disallowed, we did not fetch the page and recorded only the robots verdict. One homepage per site, at least five seconds apart, 20-second timeout, no login, no CAPTCHA bypass, no form submission and no collection of contact details.

Method change: the first 18 companies were searched on 360 Search; from company 19, that engine returned zero results for our user agent (rate limiting), so the last three used a fallback search integration with the same queries and exclusion rules, recorded per sample. Search index coverage directly affects the "no official site identified" count and is one of the largest limitations here.

Result 1: company-level discoverability

Of 21 companies, 7 had an analyzable homepage, 13 produced no identifiable official site, 1 blocked all crawlers, and 0 failed to fetch. The first finding is that being discoverable as a company website is itself a hurdle, not only a page-technical issue.

Company-level counts cover all 21. Homepage metrics in the next section describe only the 7-site sub-sample and are not converted into shares of 21.

The 13 "not identified" cases are mixed: some companies use only marketplace storefronts, social accounts or exhibition pages; some may have a new site that is not indexed; some may have no independent site. The procedure cannot distinguish these cases, so the report says "not identified", never "no website". That is also why the company-level result is an observation about search visibility, not a score of digital maturity.

Result 2: static structure of the 7 homepages

None of the 7 homepages combined a unique H1, structured data and language declarations. The most common gap was the H1: 6 had none and 1 had five. Only one had JSON-LD and none declared hreflang. For a parser that reads static HTML, these pages look visual rather than structural.

H1 and structured data are counted from static HTML only. Sites that inject headings or JSON-LD with JavaScript are undercounted; that bias is known.

Equating "no H1" with "bad SEO" would be too strong: search engines process some JavaScript, and visual titles may use other markup. Still, a unique H1 and minimal structured data are low-cost, verifiable basics, and this pilot shows the basics were not complete on these pages.

Result 3: inquiry paths and language

Of the 7 homepages, 4 had a mailto link, 2 had a tel link, 1 had a recognizable inquiry form and 1 had a privacy-policy link; none declared hreflang. Most pages show contact details, but few make the inquiry path structurally identifiable.

For buyers this means two things. First, an AI or automation tool may not reliably find a "send an inquiry" entry point on the homepage. Second, if an English or multilingual site exists, the homepage does not tell search engines where the alternate page is. We did not fetch product or inquiry pages, so we cannot infer the conversion capability of the full site.

"Recognizable inquiry form" means a static form with email/tel/textarea and a submit control; image-only or JavaScript-only forms are not counted.

Editorial judgement: what the pilot supports

It supports starting with basics, not buying a new tool. For a manufacturer, the highest-value work is to fix semantic structure, robots policy, contact paths and language alternates on the homepage, and to keep a record that can be rechecked.

1. Open the homepage as plain text. If the company introduction, products and contact details disappear without JavaScript, machines may not see them either.
2. Check the H1. One H1 per page; use CSS for visual scale instead of replacing headings with generic divs.
3. Add minimal JSON-LD. Organization (name, address, logo, sameAs) and Product/FAQ where the content genuinely supports them; never invent ratings or certifications.
4. State a robots policy. If AI search discovery matters, check whether GPTBot, OAI-SearchBot, PerplexityBot, Bytespider and ClaudeBot are blocked by accident; if they should be blocked, document why.
5. Keep the inquiry path in static HTML. Email, phone, form entry and privacy notice should not depend on JavaScript; add hreflang only if real translated pages exist.

If you are accepting a corporate website, these five items make a practical launch checklist. The AI website TCO calculator can help estimate build and maintenance cost, and the SEO and GEO guide covers crawlability and citation in more depth. Neither replaces the limits of this pilot.

Limitations: do not read this as an industry conclusion

• Specific frame: 21 companies from one AI terminal product list, not Changzhou manufacturing and not China.
• Only 7 websites: page-level metrics cannot be turned into shares, benchmarks or company rankings.
• Search-dependent: index coverage, rate limits and marketplace storefronts affect the result; the last three samples used a fallback engine.
• One fetch, one user agent, one region: TTFB is a proxy; allowing robots does not mean being crawled or cited; indexing, AI mentions, clicks and inquiries were not measured.
• Homepage only: product, inquiry and English pages were not fetched, and JavaScript was not executed.
• No freshness inference from copyright years: footer years are recorded but not used as evidence of updates.

Data and reproduction

The anonymized aggregate CSV (21 rows, sample IDs F01–F21, no company names, URLs or contact details) is available at ai-terminal-website-pilot-2026.csv. The probe is scripts/site_audit_probe.py with unit tests for robots parsing, charset sniffing and homepage metrics. Raw HTML is not published.

Recheck record: samples F04 and F11 were re-parsed from stored HTML with the same parser and all fields matched the first run. Other samples keep raw responses and timestamps for internal review. Empty cells in the CSV mean "no analyzable homepage", not zero.

Sources and definitions

1. Changzhou municipal government, "2026 Changzhou AI Terminal Innovation Product List (Second Batch)", 2026-10-08, index 014109453/2026-00054: notice.
2. RFC 9309, Robots Exclusion Protocol: rfc-editor.org.
3. schema.org Organization: schema.org.
4. Google Search Central, structured data general guidelines: developers.google.com.