Why Our Robot Directory Dropped 46,358 Records
Published: 2026-10-22 09:20 (UTC+8) | Data as of: 2026-10-22 | Source: Zhilianhai robot company directory
In October 2026 we ran a full-width/half-width normalisation pass over the English directory. The total went from 633,635 to 587,277 records: 46,358 duplicate records removed, or 7.3% of the original set. The number you see on the directory page — 587,277 — is that count.
This note explains where the duplicates came from, which copy we keep, and which figures moved when the denominator changed.
Where duplicates come from
The directory is built from public company registration data, and duplicates arrive in three shapes:
- One company, two spellings. Full-width brackets （） versus half-width (), square brackets ［］, or a full-width space inside the name. To a database these are different strings.
- Multiple registrations per group. A parent company, a sales entity and a plant entity register separately under near-identical names.
- Overlapping imports. Batches from different sources intersect; string-level matching at import time misses anything spelled differently.
The rule: normalise, then keep one
Normalisation is four steps: （）→(), ［］→[], strip full-width spaces, lowercase. Records with an identical normalised string are treated as one group.
Inside a group we keep, in order: higher data completeness score → richer record → earlier id. That means the most complete record survives; ties break to the one with more populated fields; if still tied, the earlier entry wins, so the result is reproducible.
| Step | What happens |
|---|---|
| Normalise | Full-width to half-width brackets, strip full-width spaces, lowercase |
| Group | Identical normalised strings form one group |
| Keep | score DESC, rich DESC, id ASC — one record per group |
| Delete | Remaining copies are deleted, not merged back |
What the 46,358 removed rows looked like
In a spot check, Lepu Intelligent Robot (Guangdong) Co., Ltd. existed twice under two bracket styles. After deduplication, record id 150467 was kept and the copy was deleted.
Every removal was a duplicate copy, never a company disappearing: each group keeps at least one record, and it is the most complete one. That is why the total fell by 7.3% while searching for a specific company now returns a single, cleaner entry.
The figures after deduplication
| Metric | Value | Note |
|---|---|---|
| Total records | 587,277 | was 633,635; minus 46,358 |
| Cities covered | 296 | distinct city field |
| Province-level regions | 32 | incl. municipalities |
| Records with a business summary | 383,060 | about 65.2% |
| Records with a website field | 3 | the directory does not supply websites; verify externally |
What moved, and what did not
Shares shift because the denominator changed. Industrial robots: 198,573 records, 31.3% of the old total and 33.8% of the new one. The numerator did not change; read type shares against 587,277.
Cities and provinces are counts of distinct values, so they move too. If a city had exactly one record and that record was a duplicate copy, the city drops out of the list. That is the rule working, not data loss.
Search quality improved: one company now returns one entry instead of two near-identical ones.
Limits
- Deduplication fixes “the same company counted twice”. It does not prove a company builds robots: inclusion is based on name and business-scope keywords, not certification or capability rating.
- The completeness score reflects how many fields are populated, not how strong a company is.
- Figures are as of 2026-10-22 and will change as records are added or cleaned.
Browse the deduplicated company directory by region and type, read the selection guides for how type shares should be used, or check the about page for data sources and method.