Semantic authority: what it means and how to build it

Semantic authority is an SEO term for a site that covers its subject so thoroughly and consistently that search systems can tell what the site is about. Such a site explains the subject's concepts, names the entities involved, answers the questions people ask, and ties it all together with consistent language and internal links. The name suggests a metric. No search engine publishes one, and the scores that SEO tools sell under similar names are their own estimates.

This post covers what the term means and how it relates to topical authority, what search engines and AI retrieval systems are documented to do when they judge relevance, what the scores on the market measure, and how to build the coverage. Deciding which pages a site should have at all is covered in the content and search intent guide.

The term has no agreed definition. It comes from semantic SEO, which organises content around meanings and entities. Koray Tuğberk GÜBÜR describes topical authority as a semantic SEO methodology, and says a source shows its expertise on a topic by covering more entities, attributes and sub-topics with accurate information.¹ That is close to what I mean by semantic authority: a site that does four things on its subject.

• Concepts. It explains the ideas, processes and trade-offs the subject involves, at the depth its readers need.
• Entities. It names the specific things involved, such as products, standards, organisations, places and people, and spells them the way the trade does.
• Relationships. It says how those things connect: what is a type of what, what depends on what, and what replaces what.
• Questions. It answers the questions people ask on the way to a decision, each one somewhere a reader can find it.

Consistency runs through all four. The same thing has the same name on every page, pages on related questions link to each other, and the site does not drift into subjects its readers never asked about.

How semantic authority relates to topical authority

The two terms are used loosely, so here is how I separate them. Topical authority is consistency argues that authority builds when a site covers one subject for one audience, in the same place, for a long time. That is the boundary of the lane: whether a site sticks to its subject. Semantic authority is what happens inside the lane: how completely the site covers the subject and how well its pages connect. A site can have one without the other. A focused site with ten thin pages has the first. A sprawling site with excellent coverage of one subject, buried among unrelated sections, has the second. You want both, and the work below assumes the lane is already settled.

It is also a different job from entity optimisation, which is about making your own business, people and products unambiguous. Semantic authority is about the entities of the subject you write on.

No search engine documents a semantic authority system. Several document the parts of their systems that judge meaning. Google's own explanation of relevance starts with keywords, the most basic signal that content is relevant, and a synonym system that finds pages using different words for the same thing. Its example of going further is a search for dogs: the searcher does not want a page with the word on it hundreds of times, so its systems look for other relevant content, such as pictures, videos or a list of breeds.² That list of breeds is semantic coverage at its smallest.

Google's guide to its ranking systems names several AI systems that work on meaning. BERT helps it understand how combinations of words express different meanings and intent. Neural matching helps it understand representations of concepts in queries and pages, and RankBrain helps it understand how words are related to concepts. Passage ranking identifies individual sections of a page. Google also has systems that read how pages link to each other as a way to determine what the pages are about.³ That last point is why internal linking belongs in this post.

Bing's description is close. Its relevance judgement matches terms on the page and phrases in links pointing to it, and also considers semantic equivalents, such as synonyms and abbreviations, that are understood to mean the same thing. Its judgement of quality includes the clarity of purpose of the site.⁴

Site-level signals, the leak and a patent

Most of this works page by page. Google says its ranking systems are designed to work at the page level, and that site-wide signals and classifiers also contribute to its understanding of pages.³ It does not say what those classifiers measure, which leaves two other kinds of evidence, both of which need care.

The first is the 2024 leak of Google's internal API documentation. Mike King's analysis described attributes that compare page embeddings with a site embedding to see how far a page goes outside the site's core topic, and he was clear that the documents do not show how features are weighted or whether everything in them is used.⁵ The topical authority post covers what that part of the leak does and does not show.

The second is a patent application. Google filed one in 2018 called Website representation vector. It describes a neural network turning each website's content into a vector, then classifying sites within a knowledge domain, such as health, by comparing them with the combined vectors of known groups. Its example groups are sites written by experts (doctors), apprentices (medical students) and laypeople, and it describes one site scoring differently in different domains, for relevance, authoritativeness or both. Google Patents lists the application as abandoned, a status it calls an assumption and not a legal conclusion.⁶ A patent shows what engineers explored. It is not evidence that Search uses the system, and an abandoned application is weaker evidence still.

AI assistants do not publish the details of how their retrieval works. The research, and the developer products built on the same ideas, show the general pattern. An embedding is a list of numbers that represents a piece of text, and the distance between two embeddings measures how related they are; OpenAI lists search and clustering among the uses.⁷ The dense passage retrieval paper (Karpukhin and others, EMNLP 2020) showed what that does for search: a retriever using learned embeddings beat a strong Lucene BM25 keyword system by 9 to 19 percentage points on top-20 passage retrieval accuracy, across a range of open-domain question answering datasets.⁸

Matching on exact words still matters, and retrieval products can run both kinds of search at once. Microsoft's Azure AI Search runs a full-text query and a vector query in parallel, merges the results with Reciprocal Rank Fusion, and notes that queries over product codes, highly specialised jargon, dates and people's names do better with keyword search, because it finds exact matches.⁹ Passages also lose meaning when they are cut from their page. In Anthropic's example, a chunk saying the company's revenue grew by 3% over the previous quarter does not say which company or which period. Adding that context to each chunk before indexing cut failed retrievals by 35%, and by 49% when combined with keyword matching, in Anthropic's tests, where a failure was a relevant chunk missing from the top 20 results.¹⁰

Google's AI features also search more widely than the question as asked. AI Overviews and AI Mode may use query fan-out, issuing multiple related searches across subtopics and data sources to build a response.¹¹ Google's guide to its AI features gives the example of a query about fixing a lawn full of weeds, and says its AI systems understand synonyms and general meaning, so content does not need to use the searcher's precise words. It also warns that creating separate content for every possible variation, fan-out queries included, primarily to manipulate rankings or AI responses breaks its scaled content abuse policy.¹²

Each fan-out query is, in effect, a subtopic, and each retrieval step judges passages by meaning and by exact terms. In my view that favours a site that covers the subtopics, names its entities in each passage and uses its buyers' words, because it has more passages that can match. None of this says how an assistant weighs one site against another, and no provider has documented that.

No search engine publishes a semantic authority score. The scores sold under similar names are estimates made by SEO tool companies with their own methods, and the two general authority scores in the table below are built mainly on links.

The three topic scores are the nearest thing to semantic authority, because they look at what a site covers. They are useful for comparing a site's coverage of one subject with another inside the same tool, and for judging where a new page has a realistic chance. They are the tool companies' own models of your site. A rising score means that model has changed, which may or may not show up in your rankings.

Graphite has published a study of its own score. It took 332 URLs that 12 of its client domains published in June and July 2023, split them into three equal groups by Topical Authority, and found that pages in the higher groups got their first click sooner and reached impression milestones faster. It measured early visibility only. The authors say they believe the score also affects total traffic and planned to study that separately, and the text gives no effect sizes, which appear only in charts.¹⁵ It is a vendor testing its own metric on its own clients, so read it as a sign that pages close to what a site already ranks for get picked up faster, and nothing more precise.

Most of the work is ordinary content work, done with more attention to coverage and connection. This is the order I would do it in.

Map the topic and its relationships

Start with a topic map: the subject at the top, the decisions buyers make below it, the questions below those, and one page for each intent. The content guide shows how to build one from query research. Then add the relationships. The W3C's SKOS standard, a shared model for thesauri, taxonomies and classification schemes, describes the links between concepts as hierarchical, with broader and narrower, or associative, with related.¹⁸ A topic map that records both tells you which pages should link up and down to each other and which should link across.

Cover the entities and the questions

For each page on the map, list the entities a specialist would expect it to mention and the questions a buyer would expect it to answer. The questions come from the page's queries in Search Console, the People also ask boxes on its results page, and the emails and calls the business already gets. The entities come from the subject itself: products and models, standards and certifications, regulations, organisations, places and units.

The business and its pages are invented. Each page has one job, the entities give it the vocabulary a specialist would use, and the questions become its sections. The relationships show up as links: the older-houses page links up to the installation page and across to running costs, because buyers move between the two.

To check what a finished page is mainly about, run it through an entity extraction tool. Google Cloud's Natural Language API returns the entities in a text, with a salience score from 0 to 1 for how important or central each one is to the whole document and, for most entity types, a Wikipedia URL and a Knowledge Graph ID.¹⁹ If the page's main subject does not come out as its most salient entity, the page probably spends too long on something else. The API is a Cloud product and nothing says Google Search uses the same scores, so treat it as a second reader and nothing more.

Link related pages to each other

Internal links are one of the documented ways search engines learn what pages are about. Google has systems that read how pages link to each other for that purpose,³ and Bing's relevance judgement includes the phrases used in links pointing to a page.⁴ Follow the map. Each page links up to the page on the broader subject, across to pages on related questions and down to the detail, with anchor text that says what the target covers. A new page should get links from the established pages on its parent subject on the day it goes live. The mechanics are in the internal links section of the content guide.

Use one name for each thing

Consistent terminology is the cheapest part of this. SKOS has a model for it too. Each concept gets one preferred label per language. Synonyms, near-synonyms, abbreviations and acronyms go in alternative labels, and no two concepts in one scheme should share a preferred label.²⁰

On a website, that becomes a house glossary. Pick the term your buyers use as the preferred name and use it everywhere: headings, body text, anchor text and navigation. Introduce the main alternatives once, on the page that defines the concept, as in “air source heat pump (ASHP)”. Never use one word for two things. If “service” means an annual maintenance visit on one page and the whole offer on another, readers and retrieval systems both have to guess. Google and Bing both handle synonyms,²⁴ so the aim is clarity. Exact terms still matter where the term is the fact, such as a model number or a certification, which is where keyword search does best,⁹ and a passage that names its subject still makes sense when it is lifted out of the page.¹⁰

Prune what pulls the site off its subject

Coverage also depends on what a site leaves out. Google's self-assessment for content asks whether a site has a primary purpose or focus, and lists producing lots of content on many different topics in the hope some of it performs as a warning sign.²¹ Google has also said it has systems designed to understand when a section of a site is independent or starkly different from the main content, and that it treats such sections as standalone sites, so they no longer benefit from site-wide signals. It says that is not a demotion.²² The topical authority post covers both in more detail.

Run a content audit and mark every page as on the subject, next to it, or off it. Merge the off-subject pages into something relevant, redirect them, or remove them, because they blur what the site is about and serve nobody you sell to. Do not remove pages for the sake of it: the same Google guidance says removing a lot of older content to make a site seem fresh will not help rankings.²¹ The audit process, with the questions to ask of each page, is in the content guide.

There is no score to watch, so watch what better coverage should produce.

• More queries per page. In Search Console, count the distinct queries each page in the cluster gets impressions for, quarter by quarter. A page that covers its subject should show for a widening set of related queries.
• New pages that settle faster. When you add a page inside a well-covered cluster, note how long it takes to collect impressions and how quickly its position steadies, and compare it with pages added elsewhere.
• Fewer overlaps. Two of your pages showing for the same queries means the map has an overlap to merge.
• Assistants that describe you correctly. Ask the main assistants what your site covers and who it is for, and log the answers each quarter. How to measure visibility in AI search sets out a method you can repeat.