Scraping a trademark register is the easy part
Every trademark office in Europe publishes what it receives. Free, no login, no API key worth worrying about. The Czech office even ships a daily zip of WIPO ST96 XML that curl can pull in one line.
So why is watching still a paid service? Because fetching is maybe ten percent of the job.
Quick disclosure first. I work on the marketing side at IP Defender, which sells a trademark watch. Judge the product talk accordingly. The engineering below is the same whether you buy, build, or ignore us, and every link in this post answers a plain HTTP request with a 200.
What one day of a register actually looks like
I pulled the Czech daily diff for 9 November. It came back as 213 XML records, one folder per application number.
Numbers, straight from the files. 205 records were updates, 8 were deletions. 87 were applications filed this year. The other 126 were marks from 1996, 2006, 2016 being renewed, transferred, expired or cancelled.
That is the shape of the problem in a single download. A brand new filing competing with your name sits in the same pile as a thirty year old mark that just lapsed, and nothing in the feed tells you which is which for your portfolio.
The same shape shows up everywhere else. The USPTO publishes the Trademark Official Gazette every Tuesday and pushes status changes through TSDR. EUIPO publishes Part A of the bulletin. The UKIPO has its journal. Different formats, same design. Everything is public, and none of it is addressed to you.
The clock is the product
For most pipelines a delay of a day is nothing. Here the delay is the whole value. Publication starts the opposition clock, and the clock is short.
| Office | Window to oppose | Official fee |
|---|---|---|
| EUIPO | 3 months from publication in Part A | EUR 320 |
| USPTO | 30 days from publication, extensions capped at 180 days from publication (37 CFR 2.102(c)) | USD 600 per class |
| UKIPO | 2 months, or 3 if form TM7a is filed inside the first 2 | form TM7 fee |
One detail that trips people up. The UKIPO extra month belongs to the entity that filed the TM7a. If the alert lands in a group inbox and a sibling company files the opposition, it is late. Routing matters as much as detection.
Matching is where homegrown watches die
Exact string match against the register is the version everyone writes first. It finds almost nothing that matters. Here is the order the problems bite in.
Normalization. You type Bella Vita, the register holds BELLA-VITA, or the same words with a diacritic, or a spaced out word mark. Case, spacing, punctuation, diacritics, all of it has to collapse before you compare anything.
Phonetics. Soundex and Metaphone are English tools. They do nothing useful for Czech c with a caron, Polish l with a stroke, or Hungarian gy. A mark that sounds identical to a Czech ear can score zero on an English phonetic key. You end up maintaining rules per language, and that is where most in house projects quietly stop being maintained.
Composite and reordered marks. Bella Vita against Vita Bella Cosmetics. Token overlap with order sensitivity. Plain substring matching is worse than nothing here, one shared word buries you in hits.
Goods and services. Same Nice class is not the same goods. Class 25 covers clothing and footwear, and whether two specifications actually conflict is a judgment on the wording, not on the class number.
Device marks. A large share of filings are logos. Text search misses them entirely. You need image similarity on top of Vienna codes, and Vienna coding is applied inconsistently at both ends.
Your own noise. Subsidiaries, licensees, distributors, your attorney refiling for you. If the tool pings you about your own filings in week one, you mute it in week two, and then you miss the one that mattered.
A search is a snapshot, a watch is a habit
This is the part worth repeating. A search tells you what existed on the day you ran it. Registers take new filings every working day, offices publish on their own schedule, and none of them email the owner. A few write to the applicant instead. Some publish and say nothing at all. None of them write to you.
So the money in this category is not for access to data. The data is free. It is for the permanent matching job, kept current across 55 registers, with a threshold tight enough to trust and loose enough to catch a lookalike filed in another class. That last part is a product decision as much as an algorithm, and it is the reason alert fatigue kills more watches than missed filings do.
For context on how we do it, ours runs 5 agents over 11 detection layers across 55 registers, from 6 dollars a month per country, pay as you go. The layers are described on the monitoring page, there is an honest side by side with other tools on the comparison page, and the numbers are public on the pricing page.
If you are building this yourself, start with the noise. That is what breaks first, not the fetching.
原文:https://www.a8q9.cn/en/article/scraping-a-trademark-register-is-the-easy-part-f91745.html