When someone searches "jaguar speed", a search engine has to decide whether they mean the animal or the car before it can rank anything. It makes that decision by recognising "jaguar" as one of several possible entities, using context to pick one, and then retrieving pages that are about that entity, whether or not they use the exact word "speed". That resolution step, from string to thing, is the foundation of semantic SEO. Once you understand it, a great deal of SEO advice, both old and new, sorts itself into "still true", "never true" and "true for a different reason".
This guide explains how search engines moved from keywords to entities, how they model topics, what entity salience means for a page, why covering a topic completely beats repeating a phrase, the real role of structured data, why internal links carry meaning, how to write so that passages can be extracted and cited, a practical workflow, and the misconceptions that waste time.
From keywords to entities
Early search engines indexed words. A page ranked for "payment gateway" because it contained those words, with some weighting for frequency, position and links. The system had no idea what a payment gateway was; it only knew which documents contained the string. That is why keyword density and exact-match anchors worked, and why they were easy to abuse.
Modern search engines maintain a knowledge graph: a structured model of entities (people, places, organisations, products, concepts), their attributes and their relationships. Google's version is the one behind knowledge panels. When a query comes in, the engine tries to resolve it to entities and to an intent, and when it indexes a page, it tries to identify which entities the page discusses and how central each one is. Ranking is then partly a question of how well the page's entity profile matches the query's.
Language models added another layer. Systems like the ones Google has described publicly over the years are trained on vast text and can represent the meaning of a passage as a vector, so that two passages about the same thing sit close together even when they share few words. This is why a page can rank for queries it never mentions, and why a page stuffed with the right words can fail: the model is comparing meaning, not strings.
- Query string
- Entities and intent
- Candidate pages by meaning
- Ranked by relevance, quality and context
How search engines model topics
A topic is a neighbourhood of related entities and the questions people ask about them. "UPI" as a topic includes the entity itself, the organisations around it, the apps, the transaction limits, the failure modes, the fee debates, and the common questions. A search engine can estimate, for any page and for any site, how much of that neighbourhood is covered and how deeply.
This is the mechanism behind topical authority. A site with many pages that each cover a part of the UPI topic, linked together coherently, has a stronger topic-level signal than a site with one page, even a long one. The engine is not counting pages; it is observing that the site reliably satisfies queries throughout the neighbourhood. In my experience this effect is visible in Search Console as a cluster of pages that rise together once the cluster is complete, and I set up the topical authority cluster test to measure it.
Entity salience: what a page is really about
Every page mentions many entities. Salience is the measure of how central each one is. A page about air purifiers for Delhi winters mentions Delhi, AQI, HEPA filters, specific brands and perhaps the pollution season; the salient entity should be air purifiers, with the rest as supporting context. If the page spends most of its words on the politics of Delhi's air quality, the salient entity shifts, and the page ranks for the wrong thing or nothing.
Salience is driven by position (title, H1, opening paragraph), frequency relative to other entities, and the relationships the page expresses. The practical consequence is that a page should be about one thing, say so immediately, and relate everything else to that thing. This is the semantic reason behind the old advice to "put the keyword in the title and first paragraph"; the advice survives, but the reason is clarity of subject, not string matching.
Covering a topic completely versus keyword density
Keyword density asks "how often does the phrase appear?" Semantic coverage asks "does this page address what a complete treatment of this topic would address?" The second question is the one the engine is effectively asking, because its model of the topic includes the related entities and sub-questions, and a page that handles them looks like an authoritative answer.
| Keyword-era practice | Why it worked then | Semantic equivalent | Why it works now |
|---|---|---|---|
| Repeat the target phrase several times | Term frequency was a direct ranking input | Name the subject once clearly, then cover its related entities and questions | The engine models the topic and rewards pages that span it |
| One page per keyword variant | Each string was matched separately | One page per intent, with variants absorbed naturally | Variants resolve to the same meaning; separate pages cannibalise |
| Exact-match anchor text everywhere | Anchor strings were strong relevance signals | Descriptive anchors that state what the destination is about | Anchors still inform meaning, but patterns of identical anchors look manipulated |
| Synonym lists and LSI keywords | Partially worked by widening string coverage | Entity and sub-question lists from the SERP and the knowledge graph | Coverage of related concepts is what the model actually detects |
| Long pages to maximise term matches | More words, more matches | Pages as long as complete coverage requires, structured for extraction | Depth without padding; passages can be ranked individually |
I build coverage from three sources: the questions and sub-topics visible in the SERP (People Also Ask, related searches, the headings of ranking pages), the entities a knowledge graph associates with the subject, and what an actual expert says a complete treatment needs. The third source is the one that produces pages people have not seen before, and it is where the brief earns its keep.
Structured data's real role
Structured data is often sold as the heart of semantic SEO. It is useful, but its job is narrower than that. Schema.org markup tells a search engine, unambiguously, what kind of thing a page describes and what its key attributes are: this is a product with this price; this is an article by this person; this organisation has this name, logo and the same identity as that Wikipedia entry. It removes ambiguity about identity and type. It does not make a thin page authoritative, and it does not rank pages on its own.
The parts that matter for meaning: Organisation and Person markup with sameAs links to established profiles, so your entity is connected to the graph; Article, Product and FAQ markup that mirrors visible content; and BreadcrumbList, which exposes the site's topical hierarchy. The parts that waste time: marking up things that are not on the page, and expecting ranking changes from markup alone. I measured what rich-result markup actually does in the schema markup experiment, and the technical implementation is covered in the technical SEO guide.
Internal linking as a semantic signal
Internal links are usually discussed as a way to pass authority. They are at least as important as a way to express meaning. A link from a page about UPI transaction limits to a page about UPI failure codes, with an anchor that says what the destination covers, tells the engine that these two pages belong to the same topic and how they relate. A hundred such links across a cluster define the topic's shape on your site more clearly than any single page can.
That is why I treat internal linking as an architecture problem rather than a tactic. Hubs link to spokes; spokes link back and to each other; anchors describe; links sit inside relevant sentences. The internal linking guide covers the design, and the internal linking at scale experiment looked at what changes when this is done systematically rather than by hand. The internal linking framework is the version I use on client sites.
Internal links do not just pass authority. They draw the shape of a topic, and search engines read the drawing.
Writing for extractability
Search engines increasingly rank passages, not only pages. Featured snippets, People Also Ask answers and AI-generated overviews all pull a specific section from a page and present it, with or without a click. Assistants built on language models do the same: they retrieve passages and cite them. If your page holds the right answer in a form that cannot be cleanly extracted, you lose the placement to a page that holds a worse answer in a cleaner form.
Writing for extraction is mostly good writing with a bit of discipline. Each section should answer one question, announced by its heading. The first sentence or two of a section should contain the answer, not lead up to it. Definitions should be stated as definitions. Lists and tables should be used where the content is actually a list or a table. Claims should be specific and attributable. Avoid passages that only make sense with the three paragraphs above them, because the extractor will not include those paragraphs. I set up the AI search citation test to see which of these properties actually affect whether a page gets cited, and the AI SEO guide covers the wider picture.
A practical workflow
Semantic SEO is easy to describe and hard to operationalise, because it asks you to think in topics and entities while your tools report in strings. Here is the sequence I use to bridge that gap on a real site.
Define the entities the site should own
List the core entities and topics the business is credibly about. Not every keyword cluster, but the subjects where you can be a complete source. For most businesses this is a short list; the mistake is making it long.
Map the topic neighbourhood
For each core entity, gather the related entities, attributes and questions from SERPs, knowledge-graph tools and subject experts. This becomes the cluster plan, as described in the topic clusters guide.
Assign one page per intent
Each page owns one subject and one intent. Write the subject into the title, H1 and opening lines so salience is unambiguous. Record what each page does not cover.
Write for coverage and extraction
Sections that answer stated questions, answers up front, related entities addressed in context, tables and lists where the content calls for them.
Connect the structure
Hub-and-spoke internal links with descriptive anchors, breadcrumb markup that reflects the hierarchy, Organisation and Person markup with sameAs to anchor identity.
Check and adjust
After indexing, compare the queries each page ranks for with the intent it was built for. Drift means salience is off or two pages overlap. Fix the structure before touching the words.
Misconceptions
"Semantic SEO means adding LSI keywords"
Latent semantic indexing is a decades-old technique that Google has said it does not use. The useful idea underneath, that related terms indicate topical coverage, is correct, but a list of synonyms is a weak version of it. Cover the related entities and questions because the topic requires them, not because a tool produced a list.
"Schema markup is semantic SEO"
Markup clarifies identity and type. It is one input among several and does not substitute for coverage, structure or links. A site with perfect markup and shallow content is a well-labelled shallow site.
"Google understands everything now, so structure does not matter"
Models are good at meaning and still constrained by what they can crawl, render and extract. Clear structure makes the meaning cheap to recover. Cheap-to-recover meaning gets used; expensive meaning gets skipped, especially at scale.
"Entities only matter for big brands"
Entity resolution matters most for sites whose name or subject is ambiguous, which is often small businesses. A Jaipur jewellery brand with a common name benefits more from clear Organisation markup and consistent references than a global brand that is already unambiguous.
Go deeper
- Topical authority. The site-level outcome of semantic coverage: how search engines decide a site is a credible source for a topic, and how to build towards that deliberately.
- Topical authority framework. The step-by-step process I use to turn an entity map into a cluster plan and a build order.
For the retrieval mechanics this guide assumes, see how Google Search works. For the planning layer that turns entities into pages, see content SEO.
Frequently asked questions
What is an entity in SEO?
A distinct, identifiable thing or concept: a person, organisation, place, product, event or idea that a search engine can recognise independently of the words used to name it. "Jaguar" is a string; the animal and the car company are two entities.
Is semantic SEO different from topical authority?
Semantic SEO is the practice: making meaning clear at the page and site level. Topical authority is a result: the search engine's confidence that your site is a credible source for a topic. You build the second by doing the first consistently across a cluster.
Do I still need to use my target keyword?
Yes, in the title, H1 and opening, because that is how you make the page's subject unambiguous. What you no longer need is repetition. Once the subject is clear, the rest of the page should cover the topic, not restate the phrase.
How do I find the entities related to my topic?
Read the ranking pages and note what they all cover; look at People Also Ask and related searches; check the knowledge panel and its "people also search for" area; run a sample page through a natural-language entity API; and ask someone who knows the subject what is missing. The last step is where the gaps usually are.
Does semantic SEO apply to Hindi and regional-language content?
Yes, and the entity layer is what makes multilingual search work: the engine resolves a Hindi query and an English query to the same entity. Clear subject, complete coverage and consistent markup matter in every language. Coverage of regional-language content in keyword tools is poor, so the entity-first approach is often more reliable there than string-based research.
Where this connects
The thinking here becomes a process in the topical authority framework. Its claims about extraction and citation are tested in the AI search citation test, and the architecture side is applied in the technical SEO and content architecture case study. If you want a site structured so that search engines and AI systems understand what it is about, that is the core of my AI SEO service.