Semantic Keywords and LSI: Strengthening Topical Relevance

Search engines no longer rank keyword strings. They understand topics. If you want to stay visible in DACH B2B in 2026, you have to cover a topic completely instead of placing a keyword as often as possible. This is where semantic keywords come in: topically related terms, synonyms, entities, and concepts that give a page depth. The term "LSI keywords", which has haunted SEO guides for years, is misleading. This article separates myth from substance, shows how Google actually processes language today, and lays out a practical research and optimization workflow for the German-speaking market, including AI search and EU AI Act obligations.

Why semantic keywords decide rankings today

Ranking logic has shifted from matching words to understanding meaning. It used to help to repeat a keyword like "accounting software" exactly and often. Today Google evaluates whether a page covers the topic in depth: Does it mention receipt capture, VAT ID checks, SAF-T, interfaces for FinanzOnline reporting, dunning? These associated terms are the relevance signal, not keyword frequency.

The effect can be quantified. An international analysis of 11.8 million search results (international data) shows that a higher Content Grade, meaning better semantic topic coverage as measured with Clearscope, correlates with higher rankings: an increase of 1 in Content Grade corresponds roughly to an improvement of one position (international data, top 30 results). The same study explicitly points out that word count alone does not correlate with better positions. Topical completeness decides, not length.

For B2B providers in the DACH region, this matters twice over. Services that need explaining depend on specialist terminology, and that terminology is exactly the semantic context search engines expect. If you cover a topic accurately and completely, you meet the semantic signal almost automatically.

LSI keywords: an SEO myth, put in context

Latent semantic indexing (LSI) is a technique from 1988. It mathematically reduces a term-document matrix to uncover hidden relationships between words in small, static document collections. The method predates the modern web and was never intended for a dynamic corpus of trillions of documents like Google's.

It follows that "LSI keywords" as an optimization target do not exist. Google confirms this without ambiguity. Google's John Mueller stated publicly: "There's no such thing as LSI keywords" (statement from July 30, 2019), and in 2023 he reiterated that neither LSI nor the methods marketed with it have any effect. So if you buy a tool that supposedly outputs "LSI keywords", you are buying a relabeled synonym and topic-term tool, not a piece of Google's algorithm.

What matters in practice: the concept behind it, using topically related terms naturally, remains correct and effective. Only the label and the idea of a secret keyword list are wrong. Semantic keywords, yes. "LSI keywords", no.

How Google really understands language today

Simple word matches have given way to several neural systems that capture meaning and context.

• RankBrain: The first machine learning system used in ranking interprets unseen and ambiguous queries by mapping them to known patterns with a similar meaning.
• BERT: This language model understands words in the context of the sentence rather than in isolation. Google first introduced it in 2019 for about 1 in 10 English-language searches in the US. In December 2019, Google expanded BERT to more than 70 languages, including German. Since 2020, Google says it has used BERT for almost every English-language query. BERT recognizes, for example, that the "bank" of a river is something different from a bank that holds money.
• MUM and neural matching: These systems connect concepts across languages and formats and match queries with content at the level of meaning, not characters.
• Knowledge Graph: Google's knowledge base thinks in entities, meaning clearly identifiable things such as people, organizations, places, or concepts, along with their relationships. What counts is not the string "FinanzOnline" but the entity behind it.

The consequence is clear: modern optimization targets entities and topics, not character strings. Entity-based SEO completely replaces the old LSI thinking.

Finding semantic keywords: the research workflow

Semantic research means mapping the topic field of a search intent, not ticking off a word list. The following workflow has proven itself.

• SERP analysis: The top results for your main keyword show which subtopics, terms, and entities Google expects in context. Concepts that recur across several top pages are strong semantic candidates.
• People also ask and related searches: These Google boxes surface real follow-up questions and reveal sub-intents that belong in the content.
• Entity and TF-IDF tools: Tools that compare texts against the top rankings identify terms and entities that appear more often than average in high-ranking content. They do not deliver "secret" keywords. They provide a statistical comparison of topic coverage.
• Co-occurrence instead of stuffing: The goal is for related terms to appear together naturally because the topic calls for them, not because you have a quota to fill.

The German-speaking market adds one peculiarity: compound nouns. Terms like "Lieferkettensorgfaltspflichtengesetz" (Supply Chain Due Diligence Act) or "Vorsteuerabzugsberechtigung" (entitlement to deduct input tax) pack several concepts into a single word. Good semantic research captures the compound, its components, and the related synonyms and paraphrases.

Building topical relevance and topical authority

Individual semantically optimized pages are not enough. Search engines assess how focused and coherent an entire domain's coverage of a topic field is. Topic clusters and pillar pages build this structure.

• Pillar page: A comprehensive main page covers the core topic broadly and serves as the entry point and hub.
• Cluster content: Several detail pages each answer one specific sub-question and link back to the pillar page and to each other. The result is a densely connected semantic framework instead of isolated pages.
• Internal linking: It makes the topical structure traceable for crawlers and distributes relevance within the cluster.

The 2024 Google Content Warehouse leak supports the view that Google actually measures topical focus. The documentation, a dataset of 14,014 attributes across 2,596 modules (international source), names two telling signals: siteFocusScore, which quantifies how strongly a site concentrates on one topic, and siteRadius, which measures how far a page deviates from the core topic (international source). Ahrefs also classifies these signals as the most concrete evidence so far of topical focus measured by Google (international source). One caveat matters: the leak proves that these attributes exist, not how much weight they carry in ranking. The strategic message stands: publishing with a tight topical focus beats scattering across topics.

Semantic optimization instead of keyword density

Keyword density is not a ranking factor. Forcing a specific percentage of keyword occurrences brings no advantage and produces stilted text. The question that replaces it: Does the content cover the topic completely, and does it offer something others do not?

That second point is information gain: original value beyond content that already exists. A page that only summarizes what the top results already say gains nothing semantically. A page with its own analysis, its own data, a precise DACH perspective, or a clear distinction expands the topic instead of merely covering it.

In practice, semantic optimization therefore means covering all relevant sub-questions and entities of a topic, using related terms where they belong in terms of content, and making at least one original contribution. Completeness plus original value replaces any density formula.

E-E-A-T and entities as trust signals

Semantics does not stop at the body text. Author, brand, and topical focus are entities in their own right and signal trust. Google explicitly names E-E-A-T, meaning experience, expertise, authoritativeness, and trustworthiness, as guidance and asks for people-first content, while content created primarily for search engines counts as a warning sign (official Google documentation).

Structured data makes these signals machine-readable and connects them to the Knowledge Graph.

• sameAs: Points from the Organization or Person markup to authoritative profiles such as LinkedIn, Wikidata, or an industry register and anchors brand and author as verifiable entities.
• about and mentions: State explicitly what a page is about and which entities it mentions, which reduces room for interpretation.
• Author and Organization: Link content, author, and publisher into a connected graph instead of scattering isolated snippets.

Schema does not replace trust, but it translates existing trust signals into a form that search engines and AI systems can attribute reliably. Consistency between schema and visible text is mandatory. Engines devalue contradictory information.

AI-assisted content creation: what gets rewarded and what gets penalized

AI tools have arrived in the DACH region. In Austria, 20 percent of companies with ten or more employees were already using AI technologies in 2024, compared with 11 percent in 2023 (data for Austria), nearly double within one year. The most common use is text recognition and processing at 65 percent of AI-using companies, followed by language generation at 41 percent (data for Austria). Content creation with AI is already reality, not a distant prospect.

Google's position on this is clear and nuanced. Search aims at rewarding high-quality content, however it is produced (official Google documentation). What matters is not the origin but the quality and the purpose. In the same document, Google draws the line: automation with the primary purpose of manipulating rankings violates the spam policies (official Google documentation).

In practice this means: AI may research, structure, and deliver drafts. Quality assurance stays human. Fact-checking, expert review, independent assessment, and real experience make the difference between helpful content and spam at scale. Mass-generated, interchangeable text without original value falls squarely into the penalized category.

Semantic SEO for AI search and GEO

Clear entities and documented facts directly improve visibility in AI answers. Generative engine optimization (GEO) is the discipline of getting cited in AI Overviews, ChatGPT, and Perplexity. The underlying research, published at KDD 2024, shows that GEO methods can increase visibility in generative answers by up to 40 percent (international study, measured on a benchmark of diverse queries from multiple domains).

The link to the zero-click reality is harsh. An international analysis shows that users who saw an AI summary clicked a traditional search result in only 8 percent of visits, compared with 15 percent without an AI summary (international data), and clicked a link inside the summary itself in only 1 percent of visits (international data). If you do not appear in the AI answer, you lose reach, even with a good ranking.

Semantically optimized content has the advantage here. Clearly named entities, documented statistics with sources, and precise, self-contained answer passages are exactly the building blocks an AI system can extract and cite. Topical completeness also makes a page relevant to the many sub-questions into which AI systems break down a query.

DACH and Austrian practice: language, entities, and the EU AI Act

Semantic optimization in German has its own requirements. Compound nouns bundle concepts and should be covered along with their components. In B2B, specialist terminology is not decoration. It is the semantic signal itself. Regional entities such as the WKO, FinanzOnline, the Austrian UGB, or industry-specific authorities anchor content in the right geographic and professional context.

There is also a legal dimension. Since August 2, 2026, Art. 50 of the EU AI Act has applied in the EU, with transparency obligations for AI-generated content. Providers of AI systems must mark synthetic content in a machine-readable format. For systems placed on the market before August 2, 2026, the Digital Omnibus (Regulation (EU) 2026/1744) grants a transition period until December 2, 2026. Deployers must disclose deepfakes and label AI-generated text they publish to inform the public on matters of public interest. For such text, the obligation does not apply if a human has reviewed it or exercised editorial control over it and a natural or legal person holds editorial responsibility. An AI draft that you review, revise, and take editorial responsibility for therefore needs no label under Art. 50(4). For companies that use AI in content production, this means: document the review step and the editorial responsibility, and observe the deepfake rule for images, audio, and video. These obligations do not conflict with good SEO. They match Google's call for human quality assurance and trustworthiness.

Avoiding thin content and measuring cannibalization

Semantic depth is the most effective remedy for thin content. One page that covers a topic completely and with original value replaces several thin individual pages that each serve just one term. Instead of five weak pages on five synonyms, you get one strong, comprehensive resource.

This is exactly where keyword cannibalization lurks: several of your own pages compete for the same search intent and weaken each other. Here is how you measure and avoid the problem.

• Check intent, not shared words: Two pages cannibalize each other when they serve the same search intent, not just because they share a word. Assign each page a clearly defined intent.
• Watch for SERP overlap: If different URLs of yours alternate in ranking for a keyword, or several pages sit close together in lower positions, that is a cannibalization signal.
• Consolidate instead of multiplying: Merge overlapping pages that are each weak on their own into one strong page and resolve them with redirects. This strengthens topical authority and sharpens internal linking.

In a cluster the rule is: one intent, one target page. The pillar page covers the core topic, and each cluster page covers exactly one sub-question. That keeps the semantic framework sound and the domain focused.

Next steps

Semantic keywords are neither a trick nor a secret list. They are the natural result of covering a topic completely and with expertise. "LSI keywords" are a myth. The concept of topical depth remains correct. The implementation order for DACH B2B is clear: map search intent and topic field through SERP analysis, structure content in topic clusters with pillar pages, back every core statement with a source and embed it in schema markup, anchor entities through sameAs and about, use AI only with human quality assurance and within the EU AI Act transparency rules, and track results through content coverage, topical authority signals, and AI citations. If you work this way, you optimize for classic rankings and for the citation logic of generative engines at the same time.