What is named entity recognition?
Named entity recognition (NER) is the task of finding the names of specific things in text and labeling what kind of thing each one is. Given the sentence "Tim Cook announced the new iPhone in Cupertino on September 9," a named entity recognition system marks "Tim Cook" as a person, "iPhone" as a product, "Cupertino" as a location and "September 9" as a date.
Wikipedia's definition is the standard one: NER "seeks to locate and classify named entities mentioned in unstructured text into pre-defined categories" such as people, organizations and locations [1]. The work splits into two parts: detecting where a name starts and ends, and classifying what type it is [1].
For SEO and AEO, named entity recognition matters because it is the first step in any system that reads text in terms of things rather than words. Search engines, AI assistants and content tools all have to spot the entities in a page or a query before they can connect them to a knowledge graph, judge what a page is about, or decide which source to cite.
How does named entity recognition work?
Named entity recognition works by reading text token by token and deciding, for each word or span of words, whether it is part of a name and which category it belongs to. Modern NER systems are machine learning models trained on text where people have already labeled the entities by hand.
Three families of method have been used over time [1][4]:
- Rules and lists. Early systems used hand-written patterns and gazetteers (lists of known names), such as "a capitalized word followed by Inc. is a company."
- Statistical models. From the 2000s, feature-based models such as conditional random fields learned patterns from labeled data [4].
- Neural models. Since the 2010s, recurrent networks and then transformers learn from characters, subwords and word embeddings, and generalize much better to names they have never seen [4].
The hard part is ambiguity. "Ford" can be a carmaker or a former president, "JFK" can be a person or an airport, and "the White House" often means an organization, not a building [1]. A good named entity recognition model uses the surrounding words to decide.
What types of entities does NER recognize?
Named entity recognition systems recognize a fixed set of entity types chosen by whoever built the system. The classic set is person, organization and location, which the CoNLL-2003 benchmark extended with a fourth "miscellaneous" class for names that fit none of them [3].
Google's Cloud Natural Language API uses a wider set: PERSON, LOCATION, ORGANIZATION, EVENT, WORK_OF_ART, CONSUMER_GOOD, OTHER, PHONE_NUMBER, ADDRESS, DATE, NUMBER and PRICE [6]. For phone numbers, addresses, dates and prices the API returns the parts it found, such as area code, postal code or currency [7].
The type list matters in practice. A tool that only knows person, organization and location will label a product name, a software feature or a medical procedure as "other" or miss it entirely, which says more about the tool than about the page.
What is the difference between named entity recognition and entity linking?
Named entity recognition finds a name and labels its type, while entity linking decides exactly which real-world entity that name refers to. In "Paris is the capital of France," NER tags "Paris" as a location; entity linking decides it means the French capital and not Paris Hilton, and connects it to that entry in a knowledge base [8].
Entity linking is also called named-entity disambiguation [8]. It is the step that turns a string into a thing with an ID, which is what Google meant in 2012 when it introduced the Knowledge Graph as "things, not strings" [9].
| Step | Question it answers | Output for "Apple opened a store in Paris" |
|---|---|---|
| Named entity recognition | Which words are names, and of what type? | "Apple" = organization, "Paris" = location |
| Entity linking | Which specific entity is each name? | "Apple" = Apple Inc., "Paris" = Paris, France, each with an ID |
| Entity salience | How central is each entity to the text? | A score per entity, such as Apple high, Paris lower |
Many tools run all three steps at once and report the result as "entities," which blurs the line. Google's Natural Language API, for example, returns a type for each entity (recognition), a Knowledge Graph ID and Wikipedia URL for most entities (linking), and a salience score from 0 to 1 (importance) [6].
What is the difference between named entity recognition and entity salience?
Named entity recognition decides which entities appear in a text, while entity salience measures how important each of those entities is to the text as a whole. Google's API describes salience as a score in the 0 to 1.0 range that indicates "the importance or centrality of that entity to the entire document text" [6].
A page can mention 40 entities and be about one. NER finds all 40; salience ranks them. When an SEO tool says a competitor's page "covers more entities," it is reporting recognition. When it says a page is "clearly about" a topic, it is reporting salience.
When was named entity recognition introduced?
Named entity recognition was introduced as a formal task in 1995, at the sixth Message Understanding Conference (MUC-6), a U.S. government-funded evaluation of information extraction systems [2]. Since then it has moved from hand-written rules to statistical models to neural networks, and accuracy on standard tests has risen from the high 80s to the low 90s in F1 score [3][5].
1987 to 1993: Message Understanding Conferences
The Message Understanding Conferences ran from MUC-1 in 1987, where the first corpus was ten paragraphs of naval messages, through MUC-5 in 1993 [2][4]. These early rounds tested extracting events and facts, not names on their own.
1995: MUC-6 defines the named entity task
At MUC-6 in 1995, named entity recognition was added as its own task: systems had to mark every phrase in the text as a person, location, organization, time or quantity [2]. The term "named entity" in its technical sense dates from this conference [4].
1997: MUC-7 and near-human scores
MUC-7 followed in 1997 [2]. On the MUC-7 named entity test, the best system scored an F-measure of 93.39%, against 96.95% to 97.60% for human annotators [1]. That gap is the basis for the common claim that NER on clean English news text is close to human level.
2003: CoNLL-2003 becomes the benchmark
The CoNLL-2003 shared task set the benchmark still cited today: news text in English and German, labeled with person, organization, location and miscellaneous names [3]. The best system scored 88.76 F1 on English and 72.41 on German [3], showing how much harder NER was outside English.
2018 to 2019: transformers reach Search
In October 2018, BERT was published, a language model that scored 92.8 F1 on the CoNLL-2003 English test set when fine-tuned [5]. A year later, on October 25, 2019, Google said it was using BERT in Search to understand queries in context, affecting one in 10 U.S. English searches [10]. Google did not describe BERT in Search as an NER system, but the same transformer methods now power most named entity recognition.
How do you run named entity recognition on your own page?
To run named entity recognition on a page, paste its main text into an entity analysis tool and read the list of entities, types and scores it returns. Google's Cloud Natural Language API is the common choice for SEO work because it returns Google's entity types, Knowledge Graph IDs and salience in one call [6]. Open-source libraries such as spaCy do the recognition step on your own machine.
A practical check that most guides skip:
- Strip the page to its main content. Navigation, footers and cookie banners add entities that dilute the result.
- Look for your brand and main topic in the list. If either is missing or has the wrong type, the text does not name it clearly enough.
- Check whether it is linked. An entity with a Knowledge Graph ID was matched to a known thing. A brand with a type but no ID was recognized as a name but not identified.
- Compare the top salience scores to the topic you want. If a side topic outranks the main one, the page is unclear about its subject.
One quirk of Google's documentation is worth knowing. The method description says entity analysis finds "proper names and common nouns" [7], while the Entity reference says the API currently supports proper noun mentions [6]. In practice, expect the clearest results for proper names.
Does Google use named entity recognition for rankings?
Google uses entity understanding in Search, but it has not published which named entity recognition system it uses or how entity recognition feeds rankings. What Google has said publicly is that the Knowledge Graph lets Search understand "things, not strings" [9] and that BERT helps it understand words in context [10]. Patent filings describe entity-based processing in AI search systems [11], but a patent shows what Google has invented, not what it runs.
That gap matters because of a widely repeated claim: that the Cloud Natural Language API shows "how Google sees your page." It does not. The API is a Google Cloud product for developers, and Google has never said Search uses the same model, the same entity types or the same salience scores. The API is a useful proxy for how a strong NER system reads your text. It is not a window into Google's ranking systems, and no published study shows that improving an API salience score improves rankings.
Why does named entity recognition matter for AI search?
Named entity recognition matters for AI search because answer engines have to recognize which entities a question and a source are about before they can match one to the other and cite it. When someone asks an AI assistant for "the best CRM for a five-person agency," the system has to recognize product and company names in candidate pages to compare them.
For publishers, this turns into a few plain writing habits:
- Name things fully on first mention. "Salesforce Sales Cloud," not "their CRM."
- Repeat the name instead of a pronoun in passages that may be quoted alone.
- Use one consistent form of your brand name across your site and profiles, so recognition and linking land on the same entity.
- Put the main entity in the heading and the first sentence, where it is easiest to recognize and most likely to be judged central.
These habits help any NER system, whichever one a given search engine or AI assistant uses.
When should you use named entity recognition in SEO work?
Use named entity recognition when you need to see which people, brands, products and places a piece of text actually names, rather than which keywords it repeats. It is most useful for three jobs:
- Auditing clarity. Check that a page names its main entity clearly enough to be recognized and linked.
- Comparing content. See which entities top-ranking pages name that yours does not, then judge whether those entities belong on your page.
- Checking brand consistency. Confirm that your brand is recognized as an organization, with the right ID, across your key pages.
Don't use NER output as a checklist to stuff into a page. Adding entity names a reader doesn't need makes a page less clear, and clarity is what both people and machines reward.