Your shopping cart is empty!
WordLift's new analyzer leaves an entity blank rather than guess
Publication date: 18.09.2026
In short
- WordLift has released Content Analysis v3, which turns text into stable entity identifiers in Wikidata for eight languages.
- The system abstains when the evidence is thin instead of guessing.
- On the company's own test sets it already scores above Google Cloud Natural Language for German, English and Spanish, and where it fails, the company writes about it itself.
The company that started fifteen years ago with the IKS and Apache Stanbol projects believes entity recognition is becoming basic infrastructure for AI.
So WordLift has published the third version of Content Analysis and described how it works, what it measured and where it still fails. The original would not open for us — the site returned a 403 — so the account below rests on the announcement.
From a word to a permanent identifier
The idea is easy to explain and hard to build. The word "Apple" in a text can be a fruit, a company or a surname.
Entity recognition means picking the right meaning and binding it to a stable record, in this case in Wikidata, where every object has a permanent number. Content Analysis v3 does that for eight languages and returns a set of identities other systems can point to, rather than a flat list of words.
Its developers see the task as the same kind of foundation for AI that semantic markup once was for the web.
Search engines and generative models lean on entities when they decide what a page is about and whom to trust. So it is useful for a site owner to know which entities a machine sees in their text and which it loses. We described a practical way to check that in our AI entity footprint audit checklist.
A system that knows how to abstain
The most interesting detail in the announcement is abstention. When the evidence in the text is thin, v3 leaves the entity unresolved and puts nothing in at random.
For an analyst that is the difference between a wrong link that quietly corrupts the data later and an honest gap you can see. In work where the output is loaded into a knowledge graph, that behaviour is worth more than a couple of extra points of recall.
| What is claimed | Detail |
|---|---|
| Identifiers | Stable Wikidata records |
| Languages | Eight |
| Comparison | Above Google Cloud Natural Language for German, English and Spanish on WordLift's own sets |
| Weak spots | The company describes separately where the system fails |
Why these numbers deserve caution
The comparison was run on the company's own test sets. At the time of publication there is no independent benchmark for these figures, and the announcement does not show the sample they were counted on either.
Beating Google Cloud Natural Language in three languages out of eight means exactly what it says: for the other five no such claim is made.
For our market the question is sharper: whether Ukrainian is among those eight languages is not clear from the announcement, and local companies and shops are usually thinly represented in Wikidata. If there is no record in the database, even the best tool has nothing to match against.
From that follows a simple practical conclusion about data work. Registering your brand and key products in Wikidata, together with a consistent Organization schema on the site, helps any recogniser, whether WordLift or Google. The basis of that approach is covered in our piece on a single source of truth for AI.
More than the percentages, what caught my attention is that the company talks about its own failures. That tone is rare among SEO tools, and I'd like to see whether any competitor answers in kind.
