Semantic SEO: a step-by-step guide to entities, topics and structured data

What semantic SEO means

"Semantic" means "relating to meaning". Keyword-first SEO treated a query as a string to match: pick a phrase, put it in the title, repeat it in the text. Semantic SEO starts from what the words refer to. A query names things (a product, a standard, a place, a person), asks something about them and implies a task. A page is a good result when it is about the same things and answers the same question, whether or not it repeats the exact phrase.

Google says as much in its SEO Starter Guide. Some people search for "charcuterie" and others for "cheese board", and writing with both in mind can help; but Google adds that its language matching systems can understand how a page relates to many queries even when the page never uses the exact terms. Its guide to generative AI search, last updated in July 2026, goes further: AI systems understand synonyms and general meanings, so you do not need to capture every long-tail variation.

The same guide answers the "AEO" and "GEO" question directly: from Google's perspective, optimizing for generative AI search is still SEO. Semantic SEO is best read the same way. The foundations do not change (crawlable pages, clear titles, useful content). What changes is the unit of work, from a keyword to a topic, and the target, from a matching string to a clearly described thing.

How Google learned to read meaning

Google announced the Knowledge Graph on May 16, 2012, under the title "things, not strings". Its example was the query [taj mahal]: to a keyword engine it is two words, but a person might mean the monument, a Grammy Award-winning musician or a casino in Atlantic City. The Knowledge Graph modelled real-world entities and the relationships between them, and at launch held more than 500 million objects and more than 3.5 billion facts about them and their relationships. By May 2020, Google reported over 500 billion facts about five billion entities, drawn from the web, open-source and licensed databases, and the markup content owners add to their pages.

AI systems for reading language came next. Google's ranking systems guide, last updated in December 2025, describes the ones that matter here:

BERT arrived in Search on October 25, 2019. It reads each word in the context of the words before and after it, and at launch Google said it would help with one in 10 searches in the U.S. in English. Google's example was "2019 brazil traveler to usa need a visa": earlier systems missed what the "to" meant and returned results about U.S. citizens travelling to Brazil. A year later, on October 15, 2020, Google said BERT was used in almost every English query, that passage understanding would improve 7 percent of queries across all languages, and that it was applying neural networks to understand subtopics of broad searches, such as budget, premium and small-space picks for "home exercise equipment".

MUM, announced on May 18, 2021, matters less for ranking than its launch suggested. Google described it as built on the T5 framework, 1,000 times more powerful than BERT, trained across 75 languages and multimodal, and noted that people issue eight queries on average for complex tasks. But the ranking systems guide says MUM is not currently used for general ranking, only for specific applications such as COVID-19 vaccine searches and featured snippet callouts. The same guide lists Hummingbird among its retired systems, describing it only as a major improvement to its ranking systems in August 2013.

AI Overviews and AI Mode build on the same foundation. Google says they are rooted in its core ranking systems, retrieve pages from the Search index (retrieval-augmented generation) and use query fan-out: for "how to fix a lawn that's full of weeds", related searches might include "remove weeds without chemicals" and "how to prevent weeds in lawn". A page that answers a subtopic clearly is exactly what those searches look for.

Semantic SEO vs traditional SEO

The difference is in planning and writing, not in a secret set of ranking factors:

What stays the same is everything that gets a page into the index at all: crawlable links, pages that load without errors, one canonical URL, a descriptive title. Semantic SEO sits on top of that, not instead of it.

Entities, semantic triples and the Knowledge Graph

An entity is a thing a search engine can identify: a person, organization, place, product, event or concept. Each has attributes (a founding date, a version, a location) and relationships to other entities. The Knowledge Graph stores these as a graph: entities are the nodes and relationships the connections between them.

The standard way to write such a fact is a triple. The W3C's RDF 1.1 Concepts (a Recommendation since February 25, 2014) defines the core structure as a set of triples, each made of a subject, a predicate and an object; a set of triples is an RDF graph, drawn as nodes joined by arrows. Some triples from this article:

Schema.org supplies the shared names for those subjects and predicates on the web. It was founded by Google, Microsoft, Yahoo and Yandex, its vocabulary covers entities, the relationships between them and actions, and as of version 30.1 (September 2026) it defines 826 types and 1,540 properties. JSON-LD 1.1, a W3C Recommendation since July 16, 2020, is the JSON format for writing that linked data, and the one Google recommends. When a page's JSON-LD says a BlogPosting has an author who is a Person, it states a triple: this article, has author, this person.

This is what "semantic triples SEO" means in practice, and it applies to prose as much as to markup. Write key facts as plain statements, one relationship per sentence, with the entity named in full on first mention: "BERT, a language model Google applied to Search in October 2019, reads each word in the context of the words around it" says far more than "this technology changed search". It is a clarity habit that helps readers, and any system that extracts facts, find the claim. It is not a ranking trick.

Step 1: Map the topic's entities and questions

Before writing, list what the topic is made of. Start with the main entity and the intent behind the query (what the searcher wants to know or do), then add the entities a complete answer has to mention, the attribute of each that matters, and the question each one raises. Here is the map behind this article:

The entities and questions come from four places:

1. Search Console. The Performance report's Queries tab lists what people searched when your pages appeared. For this site, that list includes comparisons ("semantic seo vs traditional seo") and narrower terms ("semantic triples seo"), each of which earned a section.
2. The results page. Google's Visual Elements gallery says you cannot control "People also ask" and related searches, but suggests paying attention to them when deciding what to write about.
3. Your own inbox. Sales calls and support tickets show which questions real customers ask, in their words.
4. Entity lookups. Google's Knowledge Graph Search API returns matching entities with schema.org types and an ID. Try an ambiguous name and see which entities and types come back, or check whether your brand or product exists as an entity at all:

curl "https://kgsearch.googleapis.com/v1/entities:search?query=BERT&limit=3&key=YOUR_API_KEY"

The API is read-only, returns individual entities rather than their connections, and Google warns it is not suitable for production-critical use (it is migrating the API to Cloud Enterprise Knowledge Graph). For a draft, Google Cloud's Natural Language API offers entity analysis, which lists the entities it detects in a text and their types (person, location, date and so on). If an entity from your map does not show up, the draft probably never names it clearly. Neither tool is part of Google Search ranking: treat their output as a check, not a score.

Step 2: Plan a topic cluster and its internal links

Turn the map into pages. A topic cluster has a hub page that covers the subject broadly and answers its main question, and cluster pages that each answer one narrower intent in depth. The hub links to every cluster page; each cluster page links back to the hub and, where it helps the reader, to its neighbours.

Split pages by intent, not by wording. Google's Query groups announcement (October 27, 2025) lists eight searches, from "how to make guacamole dip" to "guac dip recipe", that reflect a similar intent; they belong on one page. Google's AI search guide warns that creating separate pages for every variation, including the fan-out queries its AI features run, primarily to manipulate rankings violates its scaled content abuse policy. Extra pages for one intent add little visibility anyway: Google's site diversity system generally shows no more than two listings from the same site in the top results.

Then link the cluster. From Google's link best practices:

• Every page you care about should have a link from at least one other page on your site.
• Anchor text should be descriptive, reasonably concise and relevant to both pages. Cramming keywords into it is keyword stuffing.
• The words around a link matter. Do not chain several links next to each other.
• There is no ideal number of links per page.

On a site with more than a few thousand URLs, grouping each cluster in its own directory (for example /guides/tls/) also helps Google learn how often each directory changes, according to the Starter Guide. That is a crawling benefit, not a ranking one. Planning content this way is part of our SEO service: topics chosen from search demand and intent, and organized into clusters that lead back to the services they support.

Step 3: Cover each subtopic with real depth

Depth means answering the questions on your map, not adding words. Google's own guidance on helpful content asks whether a page gives a substantial, complete or comprehensive description of its topic, and whether it adds original information, reporting, research or analysis. The same page asks whether you are writing to a word count because you heard Google has a preferred one, and answers: it does not.

For each page:

1. Answer first. The first paragraph answers the main question in plain words. Passage ranking and AI features can use a section on its own, so each section should also open with its answer.
2. Name entities precisely. Use the full name and what it is on first mention ("Search Console, Google's reporting tool for site owners"), then keep the same name. Avoid pronouns that could point to two things.
3. State relationships. Say who made what, what depends on what, what replaced what, with dates.
4. Cover the subtopics a reader needs next. Your entity map is the checklist. If a subtopic deserves its own page, summarize it and link to it.
5. Use the words your readers use, including synonyms, because they arise naturally. Repetition is not coverage.
6. Add something non-commodity. Google's AI search guide favours unique expert or first-hand takes over content built on common knowledge. Your own measurements, configurations and mistakes are what nobody else can publish.
7. Show who wrote it and where the facts come from. Google's helpful content page encourages accurate authorship information, and cited sources let readers check the claims.

Step 4: Add structured data that matches the page

Google's introduction to structured data describes it as a standard way to give information about a page and classify what it contains. Google uses it to understand the page and to gather information about the world, such as the people, books or companies in the markup, and it can make a page eligible for rich results. Google supports JSON-LD, Microdata and RDFa and recommends JSON-LD; most Search features use the schema.org vocabulary, but Google's documentation, not schema.org's, is definitive for Search behaviour.

For semantic SEO, three types carry most of the weight:

Organization markup has no required properties; Google recommends adding the ones that apply and says some are used behind the scenes to disambiguate your organization from others. Article markup's author.url should point to a page that uniquely identifies the author, and Google can use both sameAs and url to disambiguate authors. Schema.org defines sameAs as the URL of a page that identifies the item beyond doubt, such as its Wikidata entry or official website, and mentions for a concept a work refers to without being about it:

<script type="application/ld+json">
  {
    "@context": "https://schema.org",
    "@type": "BlogPosting",
    "headline": "Semantic SEO: a step-by-step guide",
    "datePublished": "2024-06-19",
    "dateModified": "2026-09-25",
    "author": {
      "@type": "Person",
      "name": "Jane Example",
      "url": "https://example.com/authors/jane-example/"
    },
    "publisher": {
      "@type": "Organization",
      "name": "Example Co",
      "url": "https://example.com/",
      "sameAs": ["https://example.net/profile/example-co"]
    },
    "about": { "@type": "Thing", "name": "Search engine optimization" },
    "mentions": [
      { "@type": "SoftwareApplication", "name": "Google Search Console" }
    ]
  }
</script>

Google's structured data introduction says it can make general use of sameAs and other schema.org markup beyond the features it documents, some of which may enable future features. Treat about and mentions as clarity for anyone reading the markup, not as a lever.

Test with Google's Rich Results Test for Google features and the Schema Markup Validator for schema.org syntax in general, then watch the rich result status reports in Search Console after release, because templates break markup silently. Two limits worth knowing: structured data is not required for AI Overviews or AI Mode, and there is no special schema.org markup for them. If your templates are generated by a CMS or a JavaScript front end, emitting correct JSON-LD is often a web development change as much as an SEO one. On WordPress, see what an SEO plugin adds.

Step 5: Measure the topic in Search Console

Measure the cluster, not one keyword. The Performance report shows clicks, impressions, CTR and average position (of your topmost result), grouped by query, page, country, device, search appearance or date. Four features turn it into a topic report:

• Regex filters. In the query or page filter, choose Custom (regex) to match many variants at once. Search Console uses RE2 syntax, case-insensitive by default, and the Compare option shows the difference between two date ranges.
• Query groups. A card in Search Console Insights, announced in October 2025, that uses AI to group similar queries and shows each group's clicks. Google says the groups do not affect ranking and appear only for properties with a large volume of queries.
• Branded queries filter. Announced on November 20, 2025 and, per a March 11, 2026 update, available to all eligible sites, it splits branded from non-branded queries for top-level properties with enough query volume. Topic work shows up in the non-branded view.
• Generative AI performance report. It counts impressions of your pages in AI Overviews and AI Mode, by page, country, device and date. As of August 31, 2026, Google says it has rolled out to all websites worldwide.

A regex for this article's cluster might be:

semantic|entit(y|ies)|knowledge graph|schema markup|structured data

For structured data, Google suggests a before and after test: pick pages without markup that have several months of Search Console data and steady traffic, add the markup, confirm with the URL Inspection tool that Google found it, then compare a few months of performance filtered by URL. Give content changes the same patience; one week of data says little.

What semantic SEO does not require

A lot of advice sold as semantic SEO has no basis in Google's documentation. From the Starter Guide, the ranking systems guide and the AI search guide:

• Keyword density. Google's Starter Guide says excessively repeating the same words, even in variations, tires readers and is keyword stuffing under its spam policies.
• Word counts. Length alone does not matter for ranking; there is no minimum or maximum.
• A set number or order of headings. Semantic heading order helps screen readers, but Google says it does not matter to Search.
• E-E-A-T as a ranking factor. Google says it is not one. Its systems use a mix of signals to find content with experience, expertise, authoritativeness and trust, and give that more weight on topics that affect health, money or safety.
• llms.txt files or special AI markup. Google Search ignores them.
• "Chunking" pages for AI. Google says its systems understand several topics on one page and show the relevant part.
• A page for every query variation. As covered in step 2, it is ineffective and can count as scaled content abuse.

What does work is slower and plainer: know the topic's entities and questions, answer them better than anyone else in a well-linked cluster, describe the pages honestly in structured data, and read the results in Search Console. If you want a second pair of eyes on that, our SEO service starts from the same evidence: Search Console, server logs and field data, and every content brief mapped to one search intent and one page. More practical guides are in the SEO category.