Named entity recognition, usually shortened to NER, is the task of finding the names in a piece of text and assigning each one a type. A system reading "Vodafone moved its budget to Berlin in March" should mark Vodafone as an organisation, Berlin as a location and March as a date. The task exists because most of the world's information arrives as unstructured prose, and software cannot count, filter, link or protect what it cannot first locate. Once names are tagged, a page can be matched to an advertiser, a customer note can have its personal details masked, and an article about Apple the company can be told apart from a recipe calling for apples.

In advertising, NER rarely appears as a product. It is plumbing inside contextual targeting, brand suitability tools and privacy filters.

How a recogniser labels text

The task has two halves: finding where an entity starts and ends, and deciding what it is. Most systems treat it as sequence labelling. Text is split into tokens, and each token receives a tag. The common scheme, known as BIO notation, marks the beginning of an entity with B, tokens inside it with I, and everything else with O. In the sentence "Mark Watney visited Mars", Mark is B-PER, Watney is I-PER, visited is O and Mars is B-LOC.

The label set is a design choice rather than a law of nature. The benchmark most researchers still report against, CoNLL-2003, uses four types: person, location, organisation and a miscellaneous bucket for names fitting none of the three. The OntoNotes 5 corpus expands that to 18 tags, 11 entity types plus 7 value types such as dates and percentages, across roughly 2 million tokens. According to Google Cloud documentation, the Natural Language API's analyzeEntities method returns each entity's type, every mention in the text, and a salience score between 0 and 1 indicating how central the entity is to the whole document. In Google's own worked example, a sentence about the White House gives the entity Trump a salience of 0.79 and Washington, DC 0.03. The response can also attach a Knowledge Graph identifier, technically a second step called entity linking.

Accuracy is measured by span-based F1, the harmonic mean of precision and recall, counting a prediction as correct only when both boundaries and the type match. Tagging "Bank of" instead of "Bank of England" scores nothing.

Methods have moved through four generations. Rule sets and gazetteers, lists of known names, came first, followed by statistical models such as conditional random fields. Neural networks took over from 2016: the LSTM-CRF architecture published by Lample and colleagues that year reached 90.94 F1 on CoNLL-2003. Pretrained transformers then pushed the figure higher, with BERT Large at 92.8 in 2018 and later systems at 94.6 by 2021, according to the NLP-progress leaderboard. A fourth generation prompts a large language model (LLM) to return entities directly, trading accuracy for flexibility on new entity types.

Production systems are frequently hybrids. Microsoft's open-source Presidio toolkit, according to its documentation, combines regular expressions, NER models and rule-based logic to detect personal data, using a spaCy model by default and supporting transformer and LLM-based recognisers as alternatives.

Where it sits in advertising

The main commercial route is contextual classification. As PPC Land's explainer on the word semantic sets out, a crawler fetches a page, natural language processing (NLP) extracts entities, topics and sentiment, and the output is mapped to a taxonomy such as IAB Tech Lab's Content Taxonomy. Vendors push the resulting segments into the demand-side platform (DSP), which checks each bid request against a cached verdict for that URL. Entity extraction supplies the proper nouns that topic models tend to blur: the team, the carmaker, the politician.

The same machinery runs in reverse for brand suitability. IAS describes its Context Control product as page-level classification built on proprietary semantic technology combining NLP and what it calls cognitive analysis. Seedtag says its Liz engine combines NLP, deep learning and computer vision across more than ten languages, figures the company reports itself. Amazon's Applied Science team described a contextual system processing more than 8.5TB of webpage data a day in a technical paper released on February 11, 2025.

The third route is privacy. Before customer support transcripts, survey answers or CRM notes reach a third-party model, many organisations detect and replace personal names, places and identifiers. That detection step is NER. On the sell side, publishers encounter the technology mostly as the subject of classification; on the buy side, as a feature inside verification vendors, DSP contextual segments and internal AI tooling rather than as a setting anyone configures directly.

From MUC-6 to language models

The term came out of American defence research. According to Ralph Grishman's 2025 account in Computational Linguistics, DARPA convened a meeting in December 1993 to define goals for the Sixth Message Understanding Conference, and the committee created a Named Entity task covering people, organisations and locations. MUC-6 ran in 1995 and was the first evaluation to include it, with subtasks for entity names, temporal expressions and numerical expressions such as money and percentages. Grishman and Beth Sundheim documented that history at COLING in 1996.

The Conference on Computational Natural Language Learning (CoNLL) then made the task multilingual. The 2002 shared task covered Spanish and Dutch; the 2003 edition, organised by Erik Tjong Kim Sang and Fien De Meulder at the University of Antwerp, supplied English and German data, the English drawn from Reuters newswire. Sixteen systems took part.

Neural models, transformers and instruction-tuned LLMs each shifted the method without changing the task definition. Distillation work such as UniversalNER, published in 2023, reported a 7 billion parameter model scoring 93.30 on CoNLL-2003 after supervised fine-tuning, against 92.40 for a BERT-base baseline.

Why it matters for marketers

Cookie-free targeting depends on reading content accurately, and entity recognition decides much of what a classifier believes a page is about. Errors propagate silently into segment membership, block rates and reach. IAB Italia's AI white paper lists BERT-based named entity recognition among the NLP techniques behind trend identification and semantic discovery.

The contrast with keyword blocklists shows what is at stake. String matching cannot read context. When Vodafone narrowed its lists, the documented problem was that an entry for "die" also caught pages containing "foodie" and "diet", and the change increased its news inventory by 10 percent. Entity-aware classification is the proposed alternative, yet string matching persists: Microsoft Advertising began allowing exclusions for up to 1,000 page title terms in August 2026.

Where it fails

Benchmark scores flatter real-world performance. On the WNUT 2017 emerging-entities task, built from social media text and focused on rare and new names, the best systems listed on NLP-progress score about 50.5 F1, against more than 94 on newswire. Product launches, influencer handles and slang arrive faster than any training set.

The benchmarks themselves contain errors. A cleaned version of CoNLL-2003, known as CoNLL++, corrected about 5 percent of test instances for mislabelling, which means part of the gap between leading systems measures annotation noise rather than skill.

Prompted LLMs have not closed the gap with trained taggers. A November 2025 paper on GPT4NER reported 83.15 F1 on CoNLL-2003, roughly 88 percent of the best supervised result, and 70.37 on OntoNotes. For privacy uses, residual error is a leak. The pseudonymisation tool examined in a September 2026 study achieves recall above 91 percent, which means roughly one entity in eleven passes through unmasked.

Product changes add friction. A Google Cloud community response states that salience was removed in version 2 of the Natural Language API, so pipelines scoring entity importance on that field cannot migrate without redesign. Governance has also narrowed what text analysis can claim: the MRC's October 2025 policy bars property-level services that analyse only keywords and domains from positioning themselves as brand safety unless they examine images, video and audio too.

Not the same as

Entity linking maps a recognised mention to a specific record in a knowledge base, such as a Knowledge Graph identifier. NER says "Paris" is a location; linking decides whether it is the French capital or the city in Texas.

Keyword matching checks whether a string appears. It has no concept of type or boundary, which is the failure the brand suitability explainer describes as overblocking.

Topic classification assigns a whole page to categories such as automotive or travel. NER operates on spans within the page and often feeds the classifier.

Language detection identifies which language a text is written in, a separate step that PPC Land's language targetingexplainer distinguishes from contextual selection by subject.

Recent developments

The sharpest recent evidence concerns what happens after entities are found. A paper posted to arXiv on September 15, 2026 by researchers from the University of Bonn, Fraunhofer IAIS, the Lamarr Institute and Microsoft Germany used an NER-based pseudonymiser that replaces each name with a consistent tag. Retrieval performance for GPT-4o mini fell from 0.80 to 0.32 once entities were masked, while reasoning benchmarks typically lost under 5 points. Reversible substitution outperformed redaction on two of three benchmarks tested.

Regulation shapes how that output is treated. The EDPB's 2025 pseudonymisation guidelines hold that pseudonymised data remains personal data under the GDPR, and on July 7, 2026 the board adopted Guidelines 02/2026 on Anonymisation, with consultation open until October 30, 2026. The broader field of de-identification is therefore being redrawn at the same moment that studies put numbers on what each masking method costs. Meanwhile grounding models in retrieved documents, the task most damaged by masked entities, underpins AI search and agentic buying tools.

Timeline

  • December 1993 - DARPA meeting defines goals for MUC-6, leading to the creation of the Named Entity task.
  • 1995 - MUC-6 becomes the first evaluation to include named entity recognition, covering names, times and numerical expressions.
  • 1996 - Ralph Grishman and Beth Sundheim publish the history of MUC-6 at COLING.
  • 2002 - The CoNLL shared task covers NER in Spanish and Dutch.
  • 2003 - The CoNLL-2003 shared task introduces English and German data with four entity types; sixteen systems participate.
  • 2016 - The LSTM-CRF architecture reaches 90.94 F1 on CoNLL-2003.
  • May 2, 2017 - Google releases Cloud Natural Language API v1beta2 with expanded entity analysis languages.
  • September 19, 2017 - Entity sentiment analysis arrives in version 1 of Google's Natural Language API.
  • October 2018 - BERT Large reports 92.8 F1 on CoNLL-2003.
  • 2021 - Transformer systems reach 94.6 F1 on CoNLL-2003.
  • 2023 - UniversalNER distils LLM entity knowledge into a 7 billion parameter model.
  • February 11, 2025 - Amazon publishes a technical paper on its contextual targeting system.
  • October 18, 2025 - The MRC restricts brand safety claims by property-level verification services.
  • November 2025 - The GPT4NER paper reports 83.15 F1 on CoNLL-2003 with a prompted LLM.
  • July 7, 2026 - The EDPB adopts Guidelines 02/2026 on Anonymisation.
  • August 2026 - Microsoft Advertising rolls out page title exclusions of up to 1,000 terms.
  • September 15, 2026 - A study measuring the performance cost of NER-based pseudonymisation on five LLMs appears on arXiv.

Summary

Who. Researchers in information extraction created the task, beginning with the DARPA-funded Message Understanding Conferences and the CoNLL shared tasks. Today it is operated by cloud providers such as Google, open-source projects such as spaCy and Presidio, contextual and verification vendors, and the privacy and AI teams at advertisers and agencies.

What. Named entity recognition locates the names of people, organisations, places, dates and similar items in unstructured text and assigns each a type, usually through token-level tags scored by span-based F1.

When. The task was defined for MUC-6 in 1995, standardised for multilingual benchmarking in 2002 and 2003, and transformed by neural models from 2016 and pretrained transformers from 2018. Large language models began performing it through prompting from 2023.

Where. It runs inside contextual classification crawlers, brand suitability engines, DSP contextual segments, PII detection pipelines and the retrieval layers of AI assistants and search products.

Why. Entities anchor meaning in text. Advertising systems need them to target content without identifiers, to avoid unsuitable adjacency more precisely than keyword lists, and to strip personal data before text reaches third-party models, even though each of those uses inherits the task's error rates.