On the day we wrote this, we measured the five pages ranking for entity seo against one list of 124 subject terms. The Search Engine Land guide sitting at the top of that result uses exactly one of the 36 keyword variations on the list. It uses 28 of the 51 entities.
That ratio is the article in one line. The page winning the most competitive query in this niche is not winning it on phrasings of the target term. It is winning it on named things: Freebase, Wikidata, the Knowledge Graph, named entity recognition, Krisztian Balog, schema markup, entity disambiguation.
ProStar SEO · Entity SEO
What entities are, and what they do to a page
Things, not strings. Google, May 2012, announcing the Knowledge Graph.
The three pillars of an entity
The thing itself
People, places, organizations, products, concepts. Steve Jobs. Kyoto. NASA. The iPhone.
Its properties
Birth date, profession, address, opening hours, price. What a Knowledge Panel displays and what schema carries.
Its edges
Steve Jobs founded Apple. Apple manufactures the iPhone. Subject, predicate, object.
What the five pages ranking for “entity seo” actually place
How to read this. Bars are scaled to the largest value in the group. Every page was measured on 6 September 2026 against the same 124-term list with the same instrument, five comparable pages, same denominator rule. The list carries a small number of terms we added ourselves, so it is not a pure extraction from the results page.
Things, not strings
In May 2012 Google announced the Knowledge Graph with a phrase that has outlived most of the SEO advice written since: things, not strings.
The announcement was not marketing. Two years earlier Google had bought Freebase, a semi-structured encyclopedia where every record carried a stable machine ID. Freebase was later folded into Wikidata, and the merged foundation became the database Google consults when it tries to work out what a page is about.
Search Engine Land’s guide reports Google moving from 570 million entities to roughly 8 billion inside a decade. Treat that as reported rather than verified. Google has not published a current count, and the ones circulating in SEO articles all trace back to the same few sources.
The algorithm updates that followed all point the same direction. Hummingbird rewrote query handling around meaning instead of matching, and semantic search stopped being a research topic. RankBrain gave the system a way to handle queries it had never seen. BERT taught it to read the words around a term rather than the term alone.
None of that made keywords irrelevant. It made them insufficient.
What the retrieval side actually does with an entity
Classical keyword retrieval has a hard limit: it cannot return a document that shares no terms with the query. Everything since has been an attempt to get around that.
Krisztian Balog groups the fixes into three families. Expansion-based methods use entities to widen the query with terms the searcher did not type. Projection-based methods map the query and the document into a shared entity space and measure how close they land, which is where embeddings and vector comparisons enter the picture. Entity-based methods build an explicit entity representation of both sides and match those directly, alongside the words.
The detail worth keeping is that all three treat entities as a layer sitting over the text rather than as a replacement for it. Words still carry the signal. Entities decide what the signal means.
There is also a scale argument. No system can afford to derive the meaning of every paragraph in a corpus from scratch. Resolving mentions to known records is cheaper than reasoning about language, which is a large part of why the graph exists at all.
What are entities in SEO
An entity is a thing with an identity. In practice, four properties define it, and Krisztian Balog’s Entity-Oriented Search is the reference most of this field quotes: an entity has a name, a type, a set of attributes, and relationships to other entities. It exists, formally, when a catalog holds a record for it.
Google stores those records in the Knowledge Graph and assigns each one a machine ID, a unique identifier like /m/0k8z that never changes even when the words around it do.
That last part is what makes entities useful. Nintendo is the same entity in English, Japanese and French. The Eiffel Tower and the Torre Eiffel are one record, not two. The words change. The identity does not.
An entity catalog is a public register of those records. Several hold the ones Google leans on:
- Wikipedia, the one most people picture
- Wikidata, which stores the structured claims behind Wikipedia
- Freebase, bought by Google in 2010 and later folded into Wikidata
- Yago and DBpedia, both built on top of Wikipedia
- Your own site, once schema markup declares what you are
That last line matters more than the others for most businesses. A dental clinic in Longueuil will never have a Wikipedia article. It can still be an entity, because Schema.org exists so site owners can declare the things no encyclopedia has ever heard of.
The three pillars: the node, its attributes, its relationships
Every entity in a knowledge base is built from three parts, and each one is a place where your content either helps or stays silent.
The node is the thing itself. Entities fall into broad categories that show up on every SERP: people (Lionel Messi, Steve Jobs), places (Zion National Park, Kyoto), organizations (Google, Nike, NASA), products (the iPhone), and concepts (quantum physics, information retrieval, SEO itself). Concepts are the hardest category for Google to scale, because no single page teaches it what a vague idea means. That takes many documents, agreeing, over time.
Attributes are the properties that fill out the record. For Steve Jobs: birth date, profession, hometown, the companies he founded. For a clinic: address, opening hours, services, accepted insurance. Attributes are what a Knowledge Panel displays, and they are exactly what schema markup was designed to carry.
Relationships are the edges between nodes, and they are where meaning actually lives. Steve Jobs founded Apple. Apple manufactures the iPhone. Written out, those are subject, predicate, object:
- Steve Jobs → founded → Apple
- Apple → manufactures → iPhone
That form is not a diagram convention. It is how the graph stores facts, and it is why our writers draft body copy as explicit claims with a named subject and a checkable object. A sentence like “many businesses find that structured data helps” carries no edge. “Organization schema declares your brand’s identity to Google” carries one.
ProStar SEO · Knowledge Graph
How the graph stores a fact
Nodes are things. Edges are relationships. Attributes hang off both.
Why it matters. Entities gain meaning through edges, so a page that states relationships in plain sentences gives the graph something to read. A page that lists nouns does not.
An entity is not a keyword, and the difference is operational
A keyword is a string. An entity is what the string refers to.
| Keywords (strings) | Entities (things) | |
|---|---|---|
| What it is | A phrase typed into a search bar | A concept, person, place or object with defined traits |
| How the system holds it | A sequence of characters to match | A record with a unique identifier |
| Example | “apple company” against “apple pie recipe” | Apple Inc. against Malus domestica |
| Language | Tied to spelling and translation | Language-agnostic: Eiffel Tower and Torre Eiffel are one record |
| What it wins you | The queries you phrased for | Relevance across a neighborhood of related queries |
The practical consequence is the last row. A page built on one phrase competes for that phrase. A page built on the entities of a subject competes across every query Google considers semantically adjacent, including the ones nobody in your team thought to write down.
Disambiguation: how Google decides which Jaguar you mean
Write an article about Jaguar and the word alone tells Google nothing. What the surrounding entities say decides it.
Savanna, fur, prey, apex predator, habitat range. That is the animal. V8 engine, leather interior, dealership, model year, warranty. That is the car. Same five letters, two different records in the graph, and the page never has to say which one it means because the vocabulary around it already did.
Apple works the same way. Type it alone and the company wins, because that entity carries more weight in the index. Add orchard and the fruit wins. Mercury splits four ways between the planet, the element, the car brand and Freddie Mercury.
Ambiguity is the default state of language, and disambiguation is the process that resolves it.
Natural language processing does this at scale. Named entity recognition finds the candidate mentions in a passage of unstructured data, then the retrieval system scores which catalog entry each mention points to. Google Cloud’s Natural Language API exposes part of that with a salience score, which tells you how central it thinks each entity is to the text you paste in. TextRazor does something similar and shows you the sentences.
Neither tool is the ranking system. Both are useful as a mirror.
Where you can watch entities working
You do not need a tool to see the graph. Google shows it on the results page.
The Knowledge Panel is the entity record rendered for humans. Name, type, attributes, related entities, all pulled from the same store the ranking system consults. A Google Business Profile is the local version of the same record.
People Also Ask is a relationship map. Each question in that box is a connection Google already holds between your query and an adjacent concept, which makes it the cheapest entity research available. Related searches at the foot of the page do the same job from a different angle.
Product and image carousels are entity groupings. When Google shows a row of noise-cancelling headphones, it is not matching text. It has identified product entities and grouped them by a shared attribute.
Autocomplete leans on the same structure, which is why typing a brand name pulls its products rather than random completions.
Read those features as a free readout of what Google believes about your subject. If a concept keeps appearing in People Also Ask for your head term and appears nowhere on your page, that gap is measurable and fixable this week.
What entity based SEO actually asks you to do
Two jobs travel under the same name, and conflating them costs money.
The first job is entity coverage inside a page. A page about dental implants that names osseointegration, the abutment, the titanium post and peri-implantitis covers its subject. A page that repeats dental implants Montreal forty times does not. Coverage is what competes for non-branded queries, the ones a buyer uses before they have chosen a supplier.
The second job is brand entity identity. Organization schema on the homepage, sameAs links to real profiles, name consistency across every directory, a Google Business Profile that matches. This is what gets a Knowledge Panel, and it is what makes ChatGPT, Perplexity and Gemini describe your company correctly instead of inventing something. Brand entity seo, to give the second job its usual name, targets branded queries and citation, not acquisition.
Most articles on entity seo sell the second job while implying it does the first job’s work. It does not. A site can hold a strong brand record, rank first for its own name, and still lose the queries that introduce it to anyone new.
Distinct entities, not repeated ones
The correlation ordering we test against is the Cora shared factor table, export of 18 August 2026. Read it as an ordering rather than a measurement: the shared sample has no published size, and it is partly downstream of the tool’s own recommendations, since Cora tells every user what to change.
In that export, the top of the list is all breadth. Shared factors used sits at −0.28 to −0.31 depending on the family. Distinct entities used sits at −0.28, seventh out of roughly three thousand factors. Unique LSI terms lands next to it. Unique keyword variations follows at −0.24. Raw term frequency sits well below all of them, and exact-match density is at −0.02, which is nothing.
Repeating a term you already placed buys nothing. Placing a term you have not used yet lifts both coverage and density.
One correction to the lazy version of that finding. Exact-match repetition buying nothing does not mean the core keyword is optional. Presence is a prerequisite: a page missing the term from its title tag, its H1 and its body is very hard to rank however rich the vocabulary around it.
The reason the correlation reads near zero is narrower than it looks. Pages that omit the term are largely absent from the ranked sample being measured, so the cost of leaving it out never shows up in the number. Presence is the floor. Variation is the lever.
Placement changes what an entity is worth
The same export splits placement by category, and the split is wide enough to build headings on.
An entity in an H2 ranks at ordinal 46. An LSI term in the same H2 ranks at 467. Entities in H3 sit at 56, entities in the title tag at 72. LSI in a sentence sits at 93, which is where LSI earns its keep.
So the rule our writers follow is not stylistic. Variations lead the headings. Entities take the remaining heading slots and open the sections as sentence subjects. LSI terms live in prose. Google scores how central an entity is to a passage, and a name used as a grammatical subject reads as more central than the same name inside a parenthetical or buried in an attribute list.
Anchor text belongs in that same family. Variations in A tags correlate at −0.22, above the title tag in the current export, which makes the words you choose for internal links a placement decision rather than a formatting one.
Entity salience: which entities, and where they sit
Google’s own tooling exposes an entity salience score, which measures how central an entity is to a passage rather than whether it appears at all. The leaked documentation indicates salience scoring is used internally too. Three consequences follow, and they are the ones we build headings around.
Entity selection matters. A topically central entity supports the page; a tangentially related one dilutes it. Mentioning Shakespeare in an article about content writing feels defensible to a human, since he was a great writer with a gift for storytelling. The algorithm may read it as noise.
Placement determines weight. Entities in titles, headers and topic sentences score higher salience than the same names buried in supporting text.
Syntactic position counts. An entity used as the subject of a sentence carries more weight than one mentioned inside a list or a parenthetical aside. “Tesla revolutionized the category” and “companies such as Tesla” put the same name in the same page and do not put the same signal in it.
This is also where our honesty about the method has to sit. Semantic relevance boundaries are unknowable from outside, so some carefully chosen terms may turn out to be fluff that dilutes topical focus. Even the right entities, used in weak syntactic positions, may fail to generate the salience signal Google is looking for.
Topical authority is a cluster of entities, not a phrase repeated
Take a query like how to smoke brisket. You cannot win it by saying smoke brisket forty times.
What separates the pages that rank is whether they carry the entities the subject actually involves: the smoker, wood pellets, the rub, internal temperature, the stall, the smoke ring, resting time. A pitmaster produces that vocabulary without trying, the same way a periodontist produces osseointegration. Someone who skimmed one article produces the phrase and little else.
Topical authority, stripped of the mystique, is coverage of an entity cluster across a topic cluster: a hub page that introduces the concepts, supporting pages that go deep on each one, and internal linking that states the relationships between them.
The May 2024 Google API documentation leak surfaced internal metrics named siteFocusScore and siteRadius, which the community reads as measures of how tightly a site concentrates around a topic, with siteRadius appearing to measure deviation from the core theme. The leak also indicates Google computes embeddings for individual pages and for entire sites.
If that reading is right, entity coverage matters across the whole site rather than page by page. Scattered or off-topic content would dilute the domain’s topical identity, which is the same thing our own site-level rule has said for years: every new page should sit inside the site’s topical radius. The documents describe the existence of these metrics, not how much weight they carry, so read it as consistent with the approach rather than proof of it.
Local entity seo is the same problem with a smaller graph
For a business with an address, the entity record is mostly local. Google Business Profile is the node. Your name, address, phone, hours, categories and service area are the attributes. Citations across directories are the corroboration.
Consistency does the heavy lifting here, and it is boring work. A clinic listed as “Clinique Dentaire Rive-Sud” on its site, “Clinique Dentaire Rive Sud” on one directory and “CD Rive-Sud” on another is asking Google to merge three records it has no strong reason to merge.
The coverage half still applies. A local page that names the neighbourhoods it serves, the transit stop nearby, the procedures it performs and the insurers it accepts covers a local entity cluster. A page that repeats the city name in every heading covers nothing and reads badly to the human who has to choose a clinic.
Multi-location businesses inherit both problems at once, which is why location pages built from one template with a swapped city name underperform so consistently. Each location is a separate entity. Templates produce one entity said forty ways.
What the pages ranking for entity seo actually do
We measured five comparable pages against one merged 124-term list, using the same instrument on each. Same list, same stop-word rule, same denominator.
| Page | Words | Density | Terms used | Variations | Entities | LSI | Repetition |
|---|---|---|---|---|---|---|---|
| Search Engine Land | 6,069 | 6.6% | 56 | 1 | 28 | 27 | 4.6 |
| HubSpot | 5,502 | 9.2% | 57 | 9 | 26 | 22 | 6.3 |
| InLinks | 4,573 | 7.6% | 38 | 2 | 19 | 17 | 5.6 |
| Semrush | 2,550 | 6.2% | 33 | 2 | 18 | 13 | 3.1 |
| Clearscope | 2,221 | 7.9% | 31 | 3 | 17 | 11 | 3.8 |
Three readings come out of that table.
The entity column separates the cohort more cleanly than the variation column. Four of the five pages place between one and three keyword variations. Entity counts run from 17 to 28, and they track the ordering of the pages.
HubSpot buys its coverage lead with repetition. It reaches 57 distinct terms at a repetition index of 6.3, meaning each term it uses appears more than six times on average. Search Engine Land reaches 56 at 4.6. Same coverage, less hammering.
Length is exhaust rather than fuel. The two longest pages cover the most, because covering a subject genuinely takes words. Neither page is long for its own sake, and neither one would improve by padding.
Two caveats, because a table without them is decoration. Our list carries terms we added from our own knowledge, so it is not a pure extraction from the SERP. And the measurement ran in the browser on rendered text, which means a page without a clean content container carries some navigation into its word count.
Eighteen months of correlation data, read twice
We ran the same correlation analysis in March 2025 and again in August 2026, eighteen months apart, and published both readings side by side in our research paper. The point of running it twice was to find out whether the 2025 pattern was a snapshot or a trend.
Here is the current reading with the earlier one beside it, ordered by the 2026 coefficient.
| Ranking factor | Aug 2026 | Mar 2025 | Movement |
|---|---|---|---|
| Factors shared with the top results | −0.31 | −0.24 | stronger |
| Distinct entities used | −0.28 | −0.24 | stronger |
| Unique LSI terms used | −0.28 | −0.25 | stronger |
| Unique keyword variations used | −0.24 | −0.26 | slightly weaker |
| Variations in H1 to H3 tags | −0.23 | −0.22 | stable |
| Variations in link anchor text | −0.22 | not reported | new in 2026 |
| Word count | −0.21 | not retained | new in 2026 |
| Entities in H2 tags | −0.20 | −0.21 | stable |
| Variations in ALT attributes | −0.19 | not reported | new in 2026 |
| Entities in the title tag | −0.18 | −0.25 | weaker, largest decline |
| Number of backlinks | −0.13 (73rd) | absent from top factors | for comparison |
ProStar SEO · Correlation research
Eighteen months apart, the same ordering
Semantic ranking factors, correlation with position. Cora SEO software, published correlation studies and ProStar SEO in-house analysis. Exports of 6 March 2025 and 23 August 2026.
How to read this. Correlations run negative because a better position is a smaller number, so a longer bar means a closer link with top positions. Bars carry the August 2026 reading and the open circles mark March 2025 where the factor was measured in both exports. These are correlations across pages Google ranks, not proven causation, and Cora’s sample skews toward the US and English-language web. The durable finding is the ordering of the categories, which has held across every re-measurement since 2025.
Three of the top four factors in the 2026 reading measure how broad a page’s vocabulary is. The fourth measures how much of that vocabulary the page shares with the results already ranking. That is the finding, and it held across both readings.
The entity signals moved in opposite directions
Distinct entities used strengthened from −0.24 to −0.28 and now leads the entity family. Entities in the title tag fell from −0.25 to −0.18, the largest decline of any factor between the two exports.
Read together, those two movements say something practical. In 2025 the title tag looked like the strongest single place to put an entity. In 2026 the breadth of entities across the whole page leads instead, and title-tag entity placement sits at the bottom of the on-page factors we track.
We are not writing the title tag off. At −0.18 the relationship is weaker than we described in 2025, and it has not reversed or vanished. The core keyword belongs there regardless. What the decline argues against is the practice the old number encouraged: treating the title as a place to stack every entity that will fit. Fewer entities, better chosen, in a title that still reads as a sentence.
Why the numbers moved at all
Coefficients drifting a few hundredths between two pulls eighteen months apart is the expected result. A perfectly static reading would be the suspicious one.
Google reported 3,234 launched improvements out of 654,680 experiments in the last year for which it published complete figures. That is roughly nine changes to Search a day, almost none of them announced.
It has also said publicly that several of its core ranking systems, RankBrain and BERT among them, are machine-learned rather than hand-written. If that is right, the weighting of individual signals drifts continuously instead of stepping between fixed states. On top of that sit the confirmed broad core updates: three in 2025, four in 2024.
The sample has biases worth naming too. Cora’s dataset is built from the queries its users choose to run, which skews heavily toward the US web and English-language search. It is not a census of the internet, and no correlation tool has one.
What none of that undermines is the ordering. Across both exports, and across every re-read since 2025, the ordering of factor categories holds even as individual coefficients move: semantic variation, entity coverage and contextual density sit at or near the top every time, and raw backlink count sits far down the list every time. Treat the precise numbers as approximate. Treat the pattern as the durable finding.
Brand authority is the signal on-page work cannot supply
Entity work on a page has a ceiling, and it is worth naming where it sits.
Starting around the March 2024 core update, we saw sites with substantial branded search volume gain rankings on their non-branded keywords as well, not only on their brand terms. We hold that as a field observation rather than a measured result, because we have not isolated it from the other changes on those sites.
The outside evidence for the mechanism is more direct. The 2024 documentation leak exposed a site-level attribute named siteAuthority, applied in Google’s Q* ranking system, alongside a chromeInTotal attribute recording site-level Chrome views and the NavBoost click signals. Rand Fishkin’s first recommendation on reading those documents was to build a recognizable brand outside Google search.
The AI side points the same way. Ahrefs’ May 2025 study of 75,000 brands found branded web mentions correlating at 0.664 with visibility in AI Overviews and branded search volume at 0.392, against 0.218 for backlinks.
So brand recognition appears to act as a site-wide signal that page-level optimization cannot supply, in both classic and AI-generated results. That is the honest boundary of everything else in this article. Coverage wins you the queries your competitors are also fighting for. It does not manufacture the demand that makes people type your name.
What happened to backlinks
Backlink count sits at −0.13 in the 2026 reading, 73rd among the factors measured. In 2025 it did not appear among the top factors at all.
That is not an argument that links stopped mattering. Backlink quantity metrics move in and out of the significant factors from one pull to the next, which may reflect Google adjusting their weight or may simply be sampling variation of the kind the year-over-year drift above describes. What the data is consistent about is that raw count never rises far.
Our working hypothesis, which we have not been able to isolate in correlation data, is that the strongest links share three traits: topical relevance to the receiving site, real authority in the relevant field, and high contextual density on the linking page itself. We pay attention to the third because it is the one we can actually measure, where relevance and authority stay estimates.
Correlation tools struggle here for a structural reason. They can count links and score domains, but they cannot assess the semantic relevance and contextual density of linking pages at scale. That creates a measurement blind spot where one dense, topically aligned link might outweigh dozens of ordinary ones and still look insignificant in a correlation study.
Entities and AI search
Generative engines do not rank in the classical sense. They retrieve, then they assemble an answer, and they prefer sources they can identify with confidence.
That preference is where entity work pays a second time. A brand with a clean record, consistent attributes across its own site and public profiles, and structured data that says plainly what it sells is easier for a model to pick up and attribute. A brand described three different ways across LinkedIn, its homepage and a directory is harder, and the model either hedges or picks one version.
AI Overviews, ChatGPT, Perplexity and Gemini all lean on this in some form, whether through a live retrieval layer or through what the training corpus already contains about you. The mechanism is not mysterious. What we cannot give you is the size of the effect. Nobody has published a credible number, our own measurement of AI citation is young, and any agency quoting you a percentage lift for entity optimization in LLM answers is guessing with a straight face.
Two things are safe to say. Brand mentions without links still carry information, because a mention is an association the model can learn. And the same coverage work that wins non-branded queries also makes your page a better citation candidate, because a page that defines its terms is easier to quote than one that gestures at them.
How to find seo entities without buying another tool
Extract them. Do not invent them.
A list of entities you wrote yourself measures your own vocabulary, so your page will always score well against it. That comparison is circular, which is the single most common flaw in agency content briefs. The list has to come out of the pages Google already ranks, and out of the reference sources those pages draw on.
- Read the top five comparable pages for your query and record every named thing: brands, tools, standards, bodies, people, procedures, places.
- Open the Wikipedia article for your subject and treat its section headings as an attribute checklist. If a section exists there and nowhere on your page, you have a gap.
- Mine People Also Ask and related searches. Each one is a relationship Google already believes in.
- Run your draft through the Natural Language API or TextRazor and check that the entities coming back are the ones you meant.
- Count. A list you have not counted against your own page is a wish.
Then place what you found. Entities open sections and act as sentence subjects. Variations lead headings in volume order. LSI terms sit in prose where they read naturally, and if a term will not fit without an awkward sentence, leave it out. A forced placement reads as forced to a person, and dilution is a real risk: a tangential entity pulls the page’s center of gravity away from its subject.
A worked example: mapping one page
Take a regional dental group with four clinics, since it is the shape most of our clients are in.
The brand entity is the group’s name. It needs Organization schema, sameAs pointing at the professional order’s register, the Google Business Profile for each clinic and the real social accounts, and identical naming everywhere. That work wins branded queries and nothing else, and it is usually already half done.
The service entities are the procedures. Dental implants, root canal, crowns, wisdom tooth extraction, emergency dentistry, orthodontics. Each one deserves its own page, because each is a distinct record with its own attributes and its own demand.
The topic entities are the concepts a patient uses before they have chosen a clinic: osseointegration, bone graft, sedation options, recovery time, cost, insurance coverage. These are the ones that win non-branded queries, and they are the ones thin service pages skip entirely.
The niche entities sit under each topic: Straumann and Nobel Biocare under implants, the CBCT scan under diagnosis, the healing cap and abutment under the procedure itself. A page that names them reads as written by somebody who does the work. A page that omits them reads as written by somebody who bought a template.
Now count. Five competing pages, one merged list, one instrument, and the gaps show up as numbers rather than opinions. That is the whole method, and the vertical changes nothing about it.
Schema markup declares identity, it does not create coverage
Structured data is the cleanest signal you can send about which entity a page refers to. Organization schema with sameAs pointing at real profiles, Article schema with a named author, about and mentions carrying the subjects you actually cover. JSON-LD is the format Google reads first, and Search Console will tell you whether it parsed.
What schema will not do is make a thin page comprehensive. Clearscope’s own team says plainly that sites without schema markup still rank well when the content is entity-rich. We agree, with one addition: schema is cheap, it is verifiable, and it is the part of entity work that AI systems appear to lean on when they decide how to describe a company. Do both. Expect the coverage to carry the ranking and the markup to carry the identity.
Synonyms matter here too. An entity has aliases, and using them adds topical depth rather than repetition, because each alias is a distinct term rather than another instance of one you already placed.
Internal linking is the third leg, and anchor text is the surface that carries it. A link from your hub page to a supporting article, using the entity as its anchor, tells Google which topic cluster the two pages belong to. Consistency across those anchors compounds, and inconsistency wastes the signal.
How we count all of this
Everything above is measurable, and we measure it before a page ships. Three formulas do the work, and they are published in full in our research paper so anyone can run them.
Contextual density sums every occurrence of the subject’s vocabulary and divides by the page’s meaningful words:
CD = (Σ(KV) + Σ(RE) + Σ(LSI)) / (TW − SW) × 100
Keyword variations, relevant entities and LSI terms make up the numerator, repetitions included. Total words minus stop words make up the denominator.
Unique contextual density changes only the numerator. It counts each distinct term once, however often it recurs:
UCD = (U(KV) + U(RE) + U(LSI)) / (TW − SW) × 100
The repetition index divides one by the other:
Repetition index = CD ÷ UCD
Take a page of 1,000 useful words where subject terms appear 90 times. Its CD is 9%. If those 90 occurrences come from 30 distinct terms, its UCD is 3% and its repetition index is 3: each covered term recurs three times on average.
An index near 2 or 3 describes a page that varies its vocabulary as it goes. An index of 5 or 6 describes a page returning to a small set of terms. Two pages can post an identical CD while getting there in opposite ways, and the repetition index is what tells them apart.
UCD carries a bias of its own, in the opposite direction. Its denominator is length, so it flatters short pages. A 400-word page placing 17 distinct terms posts a higher UCD than a 4,000-word page placing 118, even though the longer page covers roughly seven times more of the subject. Neither figure is trustworthy alone.
Breadth times density, not breadth or density
Breadth is how much of the subject’s vocabulary a page uses at all, which is what UCD measures. Density is how much of the page’s useful text is made of that vocabulary. The multiplication is the point, because a page scoring near zero on either one is not saved by the other.
Broad without dense is a page that mentions everything once inside forty thousand words of filler. Dense without broad is a page repeating six terms until the percentage looks respectable. Neither is a page a reader would call good, and the product of the two is low in both cases.
Density is hard to fake in one direction and easy in the other. A page cannot reach a high CD without using the subject’s vocabulary, so padding will not get it there. Repetition will. The repetition index is the diagnostic that catches it on a report.
The entity column is the one this article is about. We count distinct entities used, where they sit, and how many of them the competing pages share. Then we set the target above the best page in the SERP rather than at some absolute number, because there is no absolute number and any agency that quotes you one invented it.
Our contextual density guide walks through the formulas with a worked example. The SEO Contextual Density Index publishes the method we use to rank real agency sites with it, and the same instrument measured every page in the table above.
This page was written to those rules and measured against them, and entity optimization is where they get applied client by client.
ProStar SEO · Mind map
Entity SEO, end to end
Six branches, in the order we work them.
The order is the method. Identity work is finite and gets maintained. Coverage never finishes, which is why the case in this article decayed two quarters after the mandate ended.
What this looks like inside a mandate
Sequencing matters more than effort, and we run it in the same order every time.
The technical floor comes first, because vocabulary cannot rescue a page Google will not fetch. Then the identity layer: one clean entity record, consistent naming, structured data that agrees with itself. That work is finite. It gets done once and then maintained.
Coverage is the part that never finishes. Each cluster gets its own term list, extracted from the pages currently ranking, and each page gets written to beat the best of them on distinct entities rather than on length. Then the page gets measured before it ships, not after somebody notices it underperformed.
Reporting closes the loop, and this is where most agency relationships quietly break. Non-branded clicks and non-branded impressions from Search Console are the headline. Branded sits beside them as context, clearly labelled, because it moves with media spend rather than with our work. Total organic traffic is the number that lets a struggling engagement look healthy for two more quarters.
One more thing we say out loud at the start. Coverage decays when nobody maintains it, because the pages you are measured against keep getting rewritten. A term list that beat the market in March describes a market that has moved by September, and a page that was comfortably ahead becomes a page that is merely present.
Where entity SEO gets oversold
Three claims in this field deserve more scepticism than they get.
The first is the entity score. Several tools will show you a number and imply Google holds something similar. Google publishes nothing of the kind, and a vendor metric that disagrees with three other vendor metrics is a house measurement, not a ranking factor.
The second is salience chasing. The Natural Language API returns a salience figure, and it is genuinely useful as a mirror for whether your page reads as being about what you meant. It is not the ranking system, it does not see your competitors, and rewriting sentences to nudge a salience decimal is time you could spend covering a concept you skipped.
The third is the Knowledge Panel as a goal. For a consumer brand with press coverage, a panel is a reasonable ambition. For a B2B supplier with forty customers, the panel will not arrive and would change little if it did. The entity record still matters, because AI systems read it, but the visible box is not the prize.
Being honest about that last one costs us proposals. It also means the ones we win are built on something we can actually deliver.
What we cannot tell you about entity SEO
Google publishes no entity score. There is no dashboard where entity authority seo appears as a number, and any agency quoting you one built it themselves.
The correlations are correlations. Pages rich in relevant entities tend to be good pages, and good pages tend to rank. We measure the pattern with confidence and we cannot prove the direction of the arrow. Our internal before and after testing carries the causal layer, on our own client base, which is not a controlled study either.
Relevance boundaries are unknowable from outside. A term that looks related to a human can register as noise, and stuffing tangential entities into a page dilutes it. More is not the goal. Right is the goal.
The correlation sample has its own bias. Cora’s dataset is built from the queries its users choose to run, which skews toward the English-language web, and coverage does not transfer evenly between markets. In our own three-market test, coverage tracked position strongly in one market and not at all in two others, because most of the term list in those two was trade vocabulary every competing page already carried.
What we exclude, and why
Some factors correlate with position without being usable as levers, either because the sample is biased or because the causation runs backwards. We report them where they appear and we do not optimize toward them.
Domain extensions correlate, because .com is over-represented in the markets measured. Load time shows paradoxes, with authority sites ranking well despite slow speeds. Social signals correlate with brand size rather than with rankings.
Word count belongs in the same category, and it is the one people argue about. It reappeared at −0.21 in the 2026 export. We report it and we do not accept it as a ranking factor.
Our reading, held as reasoning rather than proof, is that the causation runs the other way. Covering a subject properly requires a certain number of contextual terms, entities and variations, and placing those in readable sentences takes words. The word count of a strong page is the residue of its coverage.
That also explains why padding a thin page to hit a word target does nothing. The words arrive without the terms.
A metric you cannot move on purpose, or that moves as a side effect of something else, is not an optimization target.
And entity work sits on top of the technical floor, not instead of it. A page a crawler cannot fetch, render or get indexed will not be rescued by its vocabulary. Our technical SEO team clears that layer before any of this matters. E-E-A-T, brand recognition and user behaviour all carry weight that no amount of on-page vocabulary reaches.
What we publish, and what we keep
Everything in this article is reproducible. The formulas, the variable definitions, the measurement protocol and the limitations are all here, and a competent team can start measuring their own pages tomorrow without us. That is deliberate. A framework nobody else can verify is advertising rather than research.
There is a boundary, and we would rather name it than pretend it isn’t there.
Our own measurement work has surfaced a further set of signals that live in the structure of the text on top-performing pages rather than in its vocabulary: in how statements are built and related to one another, not in which terms appear. Correlation software does not surface them, because term-counting tools read what is present rather than how it is put together. Those techniques stay in-house.
Think of this article as most of the solution: enough to make entity coverage measurable, actionable and independently verifiable, and enough to move rankings on its own.
Frequently asked questions
What is entity seo in one sentence? Entity seo is the practice of optimizing a page around the recognized things a subject involves and the relationships between them, rather than around repetitions of a target phrase.
What is an entity in seo, exactly? An entity is a uniquely identifiable thing with a name, a type, attributes and relationships, stored with a stable identifier in a catalog such as the Knowledge Graph, Wikidata or Wikipedia.
How are entities different from keywords? A keyword is the string a person types. An entity is the thing that string refers to. One page can rank across many related queries because the entities it covers connect it to a whole neighborhood of meaning, not because it matched every phrasing.
What does “things, not strings” mean? It is the phrase Google used in May 2012 when it announced the Knowledge Graph. It means the system resolves your words to records in a database of real things before it decides what you want, instead of matching characters.
Do I need schema markup for entity based seo? No, and yes. Pages without schema markup rank. Schema is still the clearest way to declare which entity you are, it is cheap to implement, and it appears to matter more for how AI systems describe your brand than for classical rankings.
How do I know if my brand is an entity in the Knowledge Graph? Search your brand name and look for a Knowledge Panel. Check the Knowledge Graph Search API for an entry. Consistency across your Google Business Profile, your site and your public profiles is what builds the record over time.
How long does it take to build a brand entity? Longer than a content sprint. Entity records are built from repeated, consistent signals across sources Google trusts, so expect months rather than weeks, and expect nothing at all if the signals contradict each other.
Does entity optimization help with AI search? It appears to. Large language models retrieve and cite what they can identify confidently, and a brand with a clear entity record and consistent attributes is easier to identify. We would not present the size of that effect as a measured fact. Nobody has published a credible number.
Can you have too many entities on a page? Yes. Tangential entities dilute rather than support, because they pull the page’s apparent subject away from the one you meant. Relevance decides inclusion, never volume.
What is contextual density, and how does it relate to entities? Contextual density is the share of a page’s meaningful words drawn from its subject’s vocabulary: keyword variations, relevant entities and LSI terms, divided by total words minus stop words. Distinct entities used is one of its three components and, in the 2026 correlation reading, the strongest of them.
Will entity optimization get me to position one? No single on-page factor will. Technical foundations have to hold first, and site-level signals such as brand recognition and user engagement sit outside what page vocabulary can reach. Across 600 client sites we have seen rankings improve in roughly 80% of cases where the technical base was sound, which is an operational observation from our own client base rather than a controlled study.
Where should entities appear on the page? In the title tag, in H2 and H3 headings, and as the subjects of the sentences that open each section. Entities in headings are an order of magnitude stronger in the correlation ordering than LSI terms in the same headings.
Want to know which entities your page is missing?
We will measure one of your pages against the five ranking above it. You get the term list, the coverage gap, the repetition index, and what we would change, whether or not you work with us.