The claim being tested. Not “AI likes factories with blogs”. The narrow, checkable claim: for an assistant to name a factory, the factory has to be reachable as HTML, identify itself by legal name and location, and connect that identity to the brand a buyer would type.
The result. Five of the twelve plants serve a homepage at all. None of the eleven homepage responses fetched for this page contained a single block of structured data. None declared a language pair. Two publish a machine-readable summary, and one of those summaries describes the wrong product, because it was generated from a page that describes the wrong product.
What this is not. It is not a ranking study. Nobody outside the vendors can see inside an assistant’s index, and nothing below claims that fixing these properties will produce a recommendation. What is measured is the input side: what exists at the URL, on one day, from one network.
Six properties, in the order that matters. Each is a gate: a correct title does nothing for a page that returns 403, and a machine-readable summary does nothing if the pages it summarises are wrong.
| Layer | What it looks like in the response | Found here |
|---|---|---|
| Addressable HTML | A GET returns 200 and the sentence “we make 2–48 layer boards” is in the HTML, not painted by JavaScript after load | 5 of 12 |
| Identity in the title | The <title> names the company and the product, in the language of the page | 3 of 5 |
| Structured data | JSON-LD Organization with legalName, registration identifier, address and sameAs links | 0 of 11 |
| Language pairing | hreflang between the Chinese page and the English page, each pointing at the other | 0 of 5 |
| Entity linkage | The English page names the Chinese legal entity or its USCC, so a resolver can join the brand to the registry record | 0 of 2 |
| Machine summary | An llms.txt at the root, describing the company in plain text | 2 of 5 |
The columns are deliberately unequal because the effort is unequal. Layer 2 costs an hour. Layer 3 costs an hour per page and is the only one that states a registration number a resolver can match. Layer 6 costs an hour and is worth nothing if layer 2 is wrong — which, in this cluster, it is.
Method, stated so it can be repeated or contradicted: one GET per URL on 1 November 2026, from a single network outside China, with a desktop browser user-agent, no JavaScript executed, no cache warming. “Present” means found in the HTML of the first response. Where a site could not be reached, it is recorded as unreachable rather than as a failure of the site — a DNS record and a server config are different problems.
| Plant | Homepage | JSON-LD | hreflang | llms.txt | English page names legal entity |
|---|---|---|---|---|---|
| Yingchuangli (IQE) | HTTPS 200 | 0 | 0 | yes, both languages | no |
| Haiying | HTTPS 200 | 0 | 0 | yes (en); Chinese side serves a captcha page | no |
| Puruisen | HTTPS 200 | 0 | 0 | no | not tested |
| Kaishengshi | HTTP 200; HTTPS 403 | 0 | 0 | no | no |
| Wanyexing | HTTP 200; no HTTPS | 0 | 0 | no | no |
| Hongan Xingsheng | HTTP 404; HTTPS refused | — | — | — | — |
| Ruithai | DNS: no address | — | — | — | — |
| Zhichao (TWT) | Registered domain, no A record | — | — | — | — |
| Shengda Xinxing | No domain in its own filings | — | — | — | — |
| Jingming | No site | — | — | — | — |
| Luxshare (Suining) | Group site only, no plant page | 0 | 0 | no | no |
| Lear (Suining) | Nothing of its own | — | — | — | — |
Read as a whole: five of twelve plants serve a homepage and seven do not. Size does not predict legibility. The second-largest employer in the cluster has no site at all, the third-largest has a group-level site with no page for this location, and the largest has an English site whose title describes a different product. For seven of these twelve plants, the profile in this directory is the only English-language description of the company that exists anywhere.
The English homepage of the largest plant in the cluster — Yingchuangli, 1,706 insured employees, ¥1.359 billion of 2025 revenue, 2–48 layers and IC substrate, 146 English URLs — carries this title element:
Inno Circuits Limited-Fingerprint lock,Indoor unit
The Chinese homepage is titled correctly, with the legal name and the product line. So the error exists only on the side a foreign buyer reads, and it is not a stale template: it is the single sentence an assistant has to work with when it decides what this company sells. Nothing in the English page corrects it. The measurement is not that the site is bad — it is that the most expensive digital asset in the cluster is pointed at the wrong question.
en.iqpcb.com/llms.txt exists, 4,352 bytes of it, and it opens:
# Inno Circuits Limited Inno Circuits Limited appears to focus on fingerprint lock and indoor uni...
That is what a generated summary does. It is a mirror of the pages it was built from, so it reproduces the title error in the one file written specifically for machines. The lesson is not that llms.txt is useless; it is that llms.txt is layer six. Fixing it before fixing the title produces a faster, cleaner version of the same mistake.
Two English homepages were tested for the Chinese legal name and for the USCC. Neither appears on either page — not 四川英创力电子科技股份有限公司, not 四川海英电子科技有限公司, not an 18-character registration number. Everything a buyer would verify hangs off the Chinese name: the registry filing, the insured headcount, the certificate numbers, the customs declarations, the bank account name. The English page offers “Inno Circuits Limited” instead, which is a different legal person in a different jurisdiction.
This is the most consequential of the four, because it is invisible from inside the company. The website looks complete to the people who built it. To a resolver, the brand and the factory are two unrelated entities that happen to be in the same industry.
On both iqpcb.com and haiyingchina.com.cn, a request for a path that does not exist returned 403 Forbidden. On the Chinese side of Haiying, /llms.txt returned HTTP 200 with a 378-byte page titled “Network Security Verification”: a slider captcha, served as content.
Two different signals, both wrong. A 403 tells a crawler the resource is blocked, which invites a retry with different headers or a polite back-off; a 404 tells it the resource is absent, which is the truth. A captcha served with a 200 tells every client — machine or human — that this is what the company publishes at that address. The second is worse, because a crawler that trusts status codes will index the captcha as a description of the company.
Two of the five plants with a homepage publish an llms.txt. Yingchuangli has one in Chinese, naming the legal entity, its founding year and its location, plus an English one that repeats the fingerprint-lock positioning. Haiying has one on the English subdomain that opens with a correct, if generic, statement of what the company makes — and a Chinese domain that returns the captcha page instead.
So the cluster’s entire machine-readable output is three files, one of which is wrong, one of which is generic, and one of which does not exist. That is the honest state of “AI-ready” among twelve real factories making boards for export. It is also the reason a directory like this one has any function: it is not competing with the factories’ websites, it is supplying the sentence their websites do not contain.
For a factory that wants to be legible, cheapest first. None of the six is a purchase.
Organization. legalName, identifier (the USCC), address, sameAs pointing at the Chinese site and any registry listing. Free, one hour per page. It is the only layer that states a number a resolver can match against a filing.hreflang. Chinese page to English page and back. Free, twenty minutes. Without it, two pages about one company read as two companies.llms.txt last, generated from the finished pages rather than from a template. Free, one hour, and worthless in any other order.What not to do first. Buy a content retainer, translate the brochure into four more languages, or add a chatbot to the homepage. Each multiplies the current state rather than fixing it: four more languages of a title that names the wrong product is four more wrong titles.
Absence is not evidence. Ruithai, Shengda Xinxing, Hongan Xingsheng and Zhichao are real plants with real output — 76, 88, 86 and 1,267 insured employees respectively, the last with US$55.3 million of paid-in capital. None of the four has a reachable homepage, and the customs and registry records of all four are in this directory. If an assistant does not mention them, that is a statement about the web, not about the factory. The table above is a map of a gap, not a ranking.
Presence is not evidence either. The plant with the best digital footprint in this cluster has a homepage title about fingerprint locks, and the file it publishes for machines repeats it. A confident AI answer naming that company is not a quality signal.
What to do with both facts. Use the assistant to generate candidates, then filter them the way you would filter any list: the twenty-minute public-record check for whether the entity is real, the quotation structure for whether the offer is comparable, and the first-visit checklist for the part that cannot be done from a desk.
There is no confirmed support for the file in any major assistant, and nothing on this page depends on it. It is cheap to publish and cheap to ignore, which is why it belongs last: a summary generated from a page that describes the wrong product will describe the wrong product, and one file in this cluster does exactly that.
Not directly, and nobody outside the vendors can prove otherwise. Structured data fixes identity: the legal name, the registration number, the address, and the links that tie a brand name to one company. That matters when an answer is assembled from search results, because something has to decide whether two pages describe the same business. It does not add capability, capacity or a reason to be recommended.
Because the trading company published a page and the factory did not. In this cluster seven of twelve plants have no reachable homepage at all, while the brand names foreign buyers type usually belong to exporters and agents who do maintain pages. The assistant is not wrong about the web. It is answering a question about the web.
Yes, through registries, association lists, court and customs records, and directories like this one. A page that states a plant’s legal name, registration number, insured headcount and capacity is worth more to an assistant than a brochure site that never names the company it belongs to.
It is an argument that the first four fixes are clerical and free: a legal name on the English page, a title that names the product, structured data for the entity, and a language pair. Whether a factory then pays anyone to do the rest is a separate question. This page measures the gap rather than selling the bridge.