1. Abundant information, scarce retrieval
A normally operating mid-sized manufacturer leaves a wide range of public information online: a corporate website, storefronts on B2B platforms, job postings on recruitment sites, system and product certification records, annual reports and public disclosures, tender and supply chain directories, and exhibition participation records. Each item has value on its own, and none of them is scarce in quantity.
Yet when the goal is to retrieve "factories that meet a certain condition", these pieces of information often cannot be organised into a complete object. The retrieval system receives a set of independent texts, not a subject it can reason about.
This points to a judgement: the bottleneck in factory data is not collection. It is alignment.
2. Four kinds of misalignment
Alignment is difficult because of four structural misalignments.
Misalignment of the entity name. The name on the business licence is the full registered name. The storefront uses a brand name or abbreviation. The website homepage carries the public brand. Inside the industry there is often a customary name as well. Without the Unified Social Credit Code as an anchor, these records are several different objects at the data level.
Misalignment of industry naming. One production capability has different expressions at different points in the supply chain: what the company calls itself, what the buyer calls it, the statistical classification, and the industry research category may all differ. No single expression covers all comparable companies.
Misalignment of time. Web information carries no uniform time marker. A company that relocated three years ago and stopped production two years ago still has a website and a storefront online. The retrieval system may read old information and output an expired list.
Misalignment of credibility. The same sentence — "we operate a Class 100,000 cleanroom" — is a self-description on a website and evidence in a third-party test report. Without distinguishing sources, self-description is adopted as fact.
The combined effect of the four is counterintuitive: the more information there is, the greater the chance of failed alignment. Every additional information source adds one more opportunity for entity consolidation to fail.
3. The four-layer structure
Described the other way round, these problems define four requirements that factory data must satisfy.
The identity layer answers "which entity is this". It requires attaching records from websites, storefronts, recruitment, certifications, and annual reports to a single object. Without this layer, all subsequent work lacks a carrier.
The status layer answers "is this entity currently producing". Business registration status can only answer existing, deregistered, or revoked. It cannot answer "existing but no longer producing". This layer requires operating signals — recruitment activity, equipment purchases, supply chain transactions, exhibition records — cross-checked against one another.
The capability layer answers "what can it do, and to what standard". Product categories, process capability, equipment types, capacity scale, and certifications all belong here. It determines whether retrieval can match a real need.
The contact layer answers "who to reach, and how". Without it, the other three layers are reference material only.
The most frequently neglected of the four is the status layer. Most factory data products stop at identity and capability, because their upstream sources are business registration databases and web crawling — and "is it still producing" is answered by neither.
| Layer | Question answered | Consequence when missing |
|---|---|---|
| Identity | Which entity is this | Information cannot be consolidated; recall depends on chance |
| Status | Is it currently producing | Shell and idle companies enter the list |
| Capability | What can it do, to what standard | Cannot match a real need |
| Contact | Who to reach, and how | Results cannot enter follow-up |
4. Business registration data is not a factory database
A common misunderstanding equates a business registration database with a factory database. The difference lies in the status layer.
A registration database answers "does this company exist". As long as it is not deregistered, it stays in the database; if its scope of business mentions manufacturing or processing, it is classified as manufacturing. The count of manufacturing enterprises produced on that basis is not the same as the number of factories actually operating.
In practice, a large number of companies retain a manufacturing scope of business while engaging in trading, agency, or subcontracting, with no workshop or line of their own. For purposes such as assessing capacity, finding suppliers, or matching customers, these entities and real factories are two different kinds of object.
Tianxia Gongchang is built by layering factory identification on top of business registration information: cross-checking operating status, industry classification, and operating traces to determine whether a company is genuinely producing today, resulting in a base of 4.8 million active factories. On the surface this is only a matter of statistical definition; in practice it determines whether retrieval results are usable.
5. Two difficulties in the capability layer
Among the four layers, the capability layer's difficulty is often underestimated.
The first is category granularity. Too coarse, one term covers too much and the results contain many irrelevant companies. Too fine, users must know the formal name before they can search at all. A workable solution maintains a mapping between formal names and industry vernacular while keeping the category system detailed. Tianxia Gongchang currently maintains 1,965 sub-industry categories covering the main branches of manufacturing.
The second is divergence between description and business. A company's self-description often lags behind adjustments in its actual business. A factory's product page may still describe the categories of three years ago while it now takes on a downstream emerging application. Correcting this divergence can only be done by inferring from operating traces, not by relying on self-description.
6. Implications for factories
If retrieval depends on the four-layer structure above, then the condition for a factory to be found becomes clear: it is not about publishing more information, but about making information alignable.
Unify the entity name. Keep the company name on the website, storefront, and marketing materials as close as possible to the business licence, or at minimum state the full registered name prominently. This is the least clever measure and the most effective for machine recognition.
Describe products concretely. Generic terms have almost no discriminating power in retrieval. State the category, the process, and the downstream industry, and let both the formal name and the industry vernacular appear at least once. This measurably raises the chance of being matched.
Maintain the validity of qualifications. An expired certificate does more than affect a customer's judgement; it causes the retrieval system to exclude the company outright when filtering for candidates that hold that qualification.
Make operating traces visible. Recruitment, exhibitions, equipment purchases, and supply chain transactions are important signals of whether a company is producing. These exist objectively across multiple platforms and can be aggregated and cross-checked.
Avoid dependence on a single channel. Websites, storefronts, industry platforms, and industrial belt directories each cover a different range. Information from a single channel rarely supports a complete identity alignment.
7. Conclusion
The difficulty of factory data has long been described as a collection problem. It is in fact a structural one. Information is scattered across platforms under different names, definitions, and time markers; without a unified identity anchor, the more that is collected, the larger the surface for alignment failure.
There are two routes to a solution. One is to build the four-layer structure on the data side, completing identity, status, capability, and contact in order. The other is to improve consistency on the factory side, so that the same entity's presence across channels can be recognised as one object.
The two routes presuppose each other. The first sets the ceiling of retrieval; the second sets its floor.
Data note: the structural analysis in this article draws on the Tianxia Gongchang Industry Research Institute's ongoing review of 4.8 million active factory records, 1,965 sub-industry categories, and more than 1,000 industrial belts.