Keywords: From Strings to Entities and Semantic Understanding

Why semantic SEO determines visibility today

For years, search engines have been reading meaning, not character strings. If you optimize content for keyword density in 2026, you are optimizing for a model Google replaced back in 2013. The turning point was the Knowledge Graph: Google described it as an intelligent model that understands real-world entities and their relationships, in other words "things, not strings." At launch, it already covered more than 500 million objects and more than 3.5 billion facts and relationships between these objects. Since then, search has processed terms as linked concepts, not as isolated word sequences.

The second step was machine understanding of language. In 2015, Google introduced RankBrain, a learning system initially meant mainly to interpret queries Google had never seen before. According to Search Engine Journal, those made up about 15 percent of all search queries. Google later extended RankBrain to all queries. Semantic SEO builds on this language understanding: relevance comes from covering a topic completely, not from repeating a word.

For the B2B market in the DACH region, this matters twice over. In Austria, by far the largest share of search runs through Google: search engine market share stands at 81.87 percent for Google, 9.01 percent for Bing, and 2.75 percent for DuckDuckGo (as of May 2026). At the same time, practically all decision-makers can be reached: in January 2025, Austria had 8.69 million internet users, an online penetration of 95.3 percent. If Google understands you semantically in this market, you reach almost your entire target audience.

How meaning emerges: from string to entity and the role of BM25

An entity is a uniquely identifiable concept: a company, a product, a person, a technical term. The Knowledge Graph assigns such entities an ID and links them to attributes and other entities. For Google, "Jaguar" is not a word but either the animal, the car manufacturer, or the operating system, depending on context. This disambiguation is the core of semantic search.

Modern search systems usually work in two stages. First, a lexical method such as BM25 quickly narrows the index down to the relevant documents. Only then does semantic re-ranking with transformer embeddings order the results by meaning. The lexical stage remains the basis, and the language model refines it. Which scoring function Google uses internally is not publicly documented.

What matters is how BM25 works, because this is where the keyword density myth breaks down:

• Term frequency saturation: BM25 rewards additional repetitions of a term less and less. Beyond a certain point, the tenth occurrence adds barely any more relevance than the fifth. The parameter k1 controls this.
• Document length normalization: Longer texts contain terms more often without being more relevant as a result. The parameter b compensates for this. The default values in Elasticsearch are k1 = 1.2 and b = 0.75, and according to Elastic they work well for most corpora.

Both mechanisms lead to the same result: a higher keyword density does not improve rankings linearly. Its effect flattens out, and length normalization can even penalize it. If you pile up a keyword artificially, you optimize against the math of such methods. It makes more sense to cover the topic broadly through related terms and entities.

Entity SEO in practice: identify, link, anchor

Semantic SEO translates into concrete on-page work. The process follows three steps.

First, identify entities. Each URL gets one dominant entity, the thematic center of the page. Supporting entities frame the topic around it. A page about the "data processing agreement" should also carry entities such as GDPR, processor, and data protection authority, because these appear together in every text that covers the subject thoroughly.

Second, link entities in machine-readable form. With structured data based on Schema.org, you can declare which entity a page is about. The sameAs attribute links to authoritative references such as Wikipedia, Wikidata, or official company profiles and anchors your own entity in the Knowledge Graph. Structured data is now standard on the web: according to the Web Almanac 2024, JSON-LD appears on 41 percent of mobile pages, up from 34 percent in 2022. The most frequently marked-up types are WebSite at 12.73 percent, Organization at 7.16 percent, and LocalBusiness at 3.97 percent of mobile pages. Precise markup gives search engines and AI systems an unambiguous, verifiable data source for entity recognition.

Third, anchor the entity in the Knowledge Graph. Consistent mentions across sources stabilize an entity. Identical spelling, identical links, and a clear connection between brand, author, and topic increase the likelihood that Google treats the entity as a distinct object.

Measuring and controlling entity salience

Not every entity mentioned is equally important to a text. Salience describes how central an entity is to the entire document. Google's Natural Language API returns a salience score in the range 0 to 1.0 for this: values close to 1.0 indicate high centrality, values close to 0 indicate peripheral topics. According to the documentation, the score measures the importance or centrality of an entity to the entire document text.

You can raise salience deliberately:

• Position: The dominant entity belongs early in the text, ideally in the first sentence and in the H1. NLP analysis weights entities mentioned early more heavily.
• Frequency in moderation: The main entity should recur noticeably, but within the saturation described above, not through stuffing.
• Grammatical dependency: Entities that act as the subject or object of core statements carry more weight than mentions in subordinate clauses.

In practice, this means one URL, one dominant entity, placed consistently in the title, H1, introduction, and subheadings. If you instead spread several competing entities evenly across the text, the salience of each one drops and the topical signal gets diluted.

Best practices: semantic content architecture and on-page levers

These mechanisms result in a clear way of working that aims for topical completeness instead of keyword stuffing.

• Build topic clusters: A pillar page covers the core topic, and specialized cluster pages cover subtopics and link back. This creates a coherent topic space in which each page occupies its own dominant entity.
• Use co-occurrence: Related entities that regularly appear together in thorough, expert-level texts belong in your content. They are the signal for topical depth.
• Scope follows the topic, not the word count: Comprehensive content correlates with better rankings. The average word count of a Google top 10 result is 1,447 words, based on an analysis of 11.8 million search results. The second finding of the same analysis is revealing: keyword-stuffed title and H1 tags showed no correlation with better rankings. Completeness beats density.
• Place on-page signals consistently: The title tag, H1, and introduction carry the main entity. The heading hierarchy from H1 to H3 structures subentities logically. Internal links with descriptive anchor text that names the target entity strengthen its relevance. "Review your data processing agreement" is a more valuable anchor than "click here."

As you move from keywords to search intent, research shifts as well. Instead of match types, what counts is mapping to intent: does the user want to learn something, compare, buy, or complete a task? In the DACH B2B context, intents are often expert-level and far removed from a purchase, but they shape decisions. A page that serves the intent precisely beats one that only matches the keyword.

GEO and E-E-A-T on-page: visibility in AI answers

Generative search systems and AI Overviews are changing how content gets found. Their presence is volatile but substantial: according to a Semrush analysis of more than 10 million keywords, AI Overviews peaked at just under 25 percent of queries in July 2025 and fell back to under 16 percent by November 2025. If you want to be cited in these answers, you need Generative Engine Optimization (GEO).

The most effective GEO levers are documented. A controlled study on the GEO-bench benchmark (10,000 queries, evaluated on the test split of around 1,000 queries) shows that GEO can increase visibility in generative answers by up to 40 percent. According to the authors, the strongest individual methods include adding quotations at around 40 percent, adding statistics at around 31 percent, and citing sources at around 27 percent relative improvement on the position-adjusted word count metric, measured against the unoptimized baseline (Table 1 of the study). In concrete terms: link to trustworthy sources, back up claims with numbers, and cite statistics, just as this article does.

In parallel, E-E-A-T on-page operationalizes the trust signal. Google lists Experience, Expertise, Authoritativeness, and Trustworthiness as quality dimensions and makes clear that trust is the most important component. On-page, this means: mark up authors as their own entity, use author schema, cite sources, and add verifiable trust signals such as the update date, the legal notice, and subject-matter evidence. An author entity that is consistently linked to the topic and the brand strengthens the credibility of the entire page.

Structuring featured snippets and position zero

Structured answer formats increase your chances of winning the featured snippet. According to an Ahrefs analysis of two million featured snippets, about 12.29 percent of search queries show a featured snippet. The click distribution needs a nuanced reading, though: when the snippet appears above position 1, it receives on average only about 8.6 percent of clicks, while the result ranking directly below it gets about 19.6 percent. The snippet brings visibility and brand presence, but it does not guarantee the lion's share of clicks.

For snippet eligibility:

• Direct answer first: A concise definition or answer of 40 to 60 words directly after the question heading makes extraction easier. The same format also makes your content citable in AI answers.
• Use lists and tables: Bulleted lists and structured tables are preferred for snippets, especially for comparisons and step-by-step content.
• Questions as headings: H2 or H3 headings phrased as questions match the actual search wording and deliver the answer context directly.

Metrics and measurement: entity coverage and competitive gap

Semantic SEO can be measured, not just claimed. The relevant metrics:

• Salience score: Run your own text through Google's Natural Language API and check whether the intended main entity actually reaches the highest salience value. If an unintended entity comes out on top, the topic is not focused clearly enough.
• Entity coverage: Compare which related entities appear in thorough expert texts on the topic and which are missing from your own content. That gap is your concrete content task.
• Competitive entity gap: A semantic SERP analysis of the top results shows which entities the ranking pages cover in common. In the DACH market, this gap is often easier to close than in English-language competition, because fewer in-depth German-language sources compete.
• Structured data validation: Use the Schema Markup Validator and the Rich Results Test to make sure that entity markup and sameAs links are read without errors.
• Visibility in rich results and AI Overviews: Track whether pages appear as rich results and whether they are cited in generative answers. The latter is the direct indicator of successful GEO work.

Further reading

Semantic SEO is not an add-on to on-page optimization but its framework. If you define one dominant entity per URL, anchor it with Schema.org and sameAs, control salience through position and structure, and cover the topic completely through related entities, you optimize for classic rankings, featured snippets, and AI answers at the same time. Sensible next steps are studying on-page schema markup in more depth, operationalizing E-E-A-T through author entities, and connecting your internal links into a consistent topic cluster architecture. The common denominator remains: meaning beats strings.