Topical relevance is something you can count, and most pages about it never try

Topical relevance decides whether a search engine reads your page as being about its subject or merely mentioning it. Every guide to it says the same three things: cover your niche, build topic clusters around relevant topics, earn relevant links. None of them says what a search engine would actually count, which makes it advice you cannot check. We measure relevance on client pages before we write and after we publish, and this is the method, the arithmetic, and the part we cannot prove.

What topical relevance actually means

Topical relevance is the degree to which a page’s vocabulary matches the vocabulary a search engine associates with a subject. That definition is deliberately narrow. It says nothing about quality, nothing about trust, and nothing about how long you have been publishing. It describes one thing: whether the words on the page place it inside the subject or outside it.

A page about oculoplastic surgery that uses the words blepharoplasty, ptosis, orbit, canthus, levator muscle and post-operative edema sits inside the subject. A page that says “eyelid surgery” nine times and nothing else sits outside it, and it sits outside it no matter how many times it repeats the phrase. Repetition does not move a page toward a subject. Vocabulary does.

Several names circulate for roughly the same idea. Topical relevance, topical relevancy, content relevance, semantic relevance, keyword relevance and topical SEO are used more or less interchangeably across the industry, and semantic SEO is the umbrella most people file them under. The terms differ in scope rather than in kind. Keyword relevance asks whether a page matches one query. Topical relevance in SEO asks whether it matches a subject, which is a set of related terms and related topics rather than a string. Nobody publishes a relevance score you can look up, so the relevance signals you can act on are the ones you can count yourself.

Topical relevance is a property of a page, not a promise about your niche

Most writing on this subject slides between two different claims without noticing. The first claim is that a page is relevant to a query. The second is that a site is relevant to a topic. They are different measurements with different inputs, and conflating them is why so much advice about topical relevance is unactionable.

Page relevance is countable today. You can open a page, extract the terms a subject requires, and count how many of them appear and how often. Site relevance is not countable by you, because it depends on how a search engine has embedded your whole domain, and you cannot see that. Site-level topical coverage is still something you can influence through what you choose to publish, which is not the same as being able to measure it.

Topical relevance vs topical authority

Topical relevance is a measurement. Topical authority is a reputation.

Relevance answers a question about a single page: does this text belong to this subject? Authority answers a question about a site: is this source the one to cite on this subject? Relevance can be established by one well-built page in an afternoon. Authority accumulates over months from coverage, citations, brand searches and links, and no single page creates it.

The practical consequence is that they fail differently. A relevant page on a site with no authority will rank for specific, low-competition phrasings and stall on the head term. An authoritative site can rank a thin page on a head term, which is exactly what Ahrefs does with a 359-word glossary entry sitting at position two for this very keyword. Its text carries almost no topical vocabulary. Its domain carries everything.

So relevance is necessary and insufficient. That is an uncomfortable thing for an agency to publish, and it is true.

Why the distinction changes what you do on Monday

If you believe the two are the same thing, you write more pages and wait. If you know they are different, you run two jobs: make each page measurably relevant, which you control and can verify this week, and build topical authority through coverage, citations and brand, which is slow and only partly yours. Teams doing only the second publish a lot of vaguely on-topic content and wonder why their rankings never move.

How a search engine could be computing relevance

We cannot observe Google’s ranking systems. We can observe what its own documents describe, what its published research says, and which measurable properties of a page correlate with position. That is the whole evidence base, and anyone claiming more than that is selling something.

Embeddings turn a page into a position, not a keyword list

Modern retrieval systems represent text as vectors. A model reads a document and outputs a list of numbers that locates it in a high-dimensional space where documents about similar things end up near each other. Google has described this approach in its own research for years. Hummingbird began the shift away from string matching in 2013, and BERT, RankBrain and MUM all operate on representations of this kind. The helpful content update pushed the same logic to site level.

What matters for relevance is what falls out of that design. If a page is a position in a space, then the words that put it there are the ones that carry meaning, and repeating one of them does not move the position much. Adding a word the subject requires and the page lacks moves it a lot. Vocabulary breadth changes the coordinates. Frequency mostly does not.

That is a mechanism, not a proof. It does explain why every correlation study run on organic search for the past decade finds the same shape.

What the 2024 leak named: siteFocusScore and siteRadius

In May 2024 a large body of internal Google API documentation became public. Two attribute names in it are directly about the subject of this page.

siteFocusScore is described by the community analysis as a measure of how concentrated a site’s content is on a single topic. siteRadius is described as how far an individual page’s content sits from that site’s central theme. Alongside them sits site2vecEmbedding, which holds vector embeddings of an entire site rather than a page.

Read together, those three names describe a system that embeds pages, embeds sites, and measures the distance between the two. That is site-level topical relevance, computed.

Independently of the leak, Google holds a patent on site-wide quality scoring (US9031929B1), and has said publicly that several of its core ranking systems are machine-learned rather than hand-written. None of that tells us the weight any of these carries.

What the leak does not say, and what we refuse to invent

The documentation names attributes. It does not say how they are calculated, whether they are used in live ranking, how heavily they are weighted, or whether the values have changed since the documents were written. Every confident article you have read explaining “how siteFocusScore works” is extrapolating.

The names tell us the shape of the problem Google is solving and nothing about the magnitude, so we treat the Google API leak as confirmation that topical concentration is measured in some form and we do not build a score on it.

One correction to the common reading: a high focus score is not a quality score. The New York Times is a high-quality site with almost no topical concentration. Focus and quality are separate assessments, and a site can be excellent and unfocused, or narrow and worthless.

The measurement gap in every page ranking for this keyword

We measured the five comparable pages currently ranking for “topical relevance” against one merged 151-term list, using the same script on every page.

Across all five, these terms appear zero times: embeddings, vector, siteFocusScore, siteRadius, query fan-out, AI Overviews, ChatGPT, Perplexity, knowledge graph, E-E-A-T, BERT, RankBrain, correlation, benchmark, breadth, distinct terms, measurable.

Every one of those pages asserts that search engines assess relevance. Not one explains what a search engine could be counting, and not one offers a way to check whether a given page has any. The recommendations are always the same: write in-depth content, cluster your topics, get relevant links, watch your bounce rate.

The measurements behind that paragraph. The positions are a Semrush reading taken in September 2026 rather than a live check, so treat them as the state of the SERP on that date:

Density runs backwards against position here. The two densest pages in the set rank eighth and tenth. The page with the widest vocabulary ranks tenth. On this particular SERP, text metrics are not what separates the winners, and we would rather say that than pretend our own numbers explain a result they do not explain. What separates them is domain strength, which is why a 359-word glossary stub holds position two.

That makes the measurement a floor rather than a lever. Being the thin page in the search engine results pages is a way to lose. Being the dense one is not by itself a way to win.

Contextual density: relevance expressed as a number

We needed a number for relevance because you cannot compare what you cannot count. Contextual density is what ProStar SEO built, and the full derivation lives in our page on contextual density.

The formula

Contextual Density (CD) = (Σ(KV) + Σ(RE) + Σ(LSI)) / (TW − SW) × 100

• Σ(KV) is every occurrence of a keyword variation: morphological, syntactic and semantic variants of the target term.
• Σ(RE) is every occurrence of a relevant entity: the proper nouns, organizations, products, standards and named concepts the subject involves.
• Σ(LSI) is every occurrence of a contextually co-occurring term: the vocabulary that reliably surrounds the subject without being a variant of it.
• TW is the total word count of the body text, excluding markup and metadata.
• SW is the stop word count, the function words carrying minimal semantic load.

The result is the percentage of meaningful words on the page that are doing topical work. It is not keyword density. Keyword density counts one string. Contextual density counts the subject’s whole vocabulary, and the exact keyword is a small fraction of it.

The three categories a term list is built from

Each of the three inputs answers a different question about coverage, and a list missing one of them measures the page badly.

Keyword variations are the ways the subject itself can be said. Plural and singular, verb and noun, the abbreviation and the spelled-out form, and the long-tail extensions people actually type. “Topical relevance”, “topically relevant content”, “topical relevance for SEO” and “topic relevance” are four variations of one idea, and a page using all four covers more search queries than a page repeating one of them four times.

Named entities are the things the subject involves: organizations, products, tools, standards, people, places. Google resolves these against its knowledge graph rather than treating them as strings, and entity coverage across the whole page is the strongest placement-independent signal in our 2026 correlation reading. A page about search engine optimization that never names Google, BERT, the knowledge graph or an analytics tool has not covered its entities, whatever its topical focus claims to be.

LSI terms are the vocabulary that reliably surrounds the subject without being a variant of it or an entity in it. The name is a historical accident, since latent semantic indexing is a specific 1980s technique and not what modern retrieval does, but the category it points at is real: the co-occurring words a genuine subject-matter expert uses without thinking about it. Search intent, anchor text, crawl budget, thin content and content gap are LSI terms for this page. Subtopics, related topics and semantically related phrasings all land in the same bucket.

A term list is the union of the three, extracted from the pages already ranking. It is never written from your own draft, because a list written from your draft will always score your draft well.

Unique contextual density, because occurrences lie

CD sums occurrences, and that is a real weakness: a term used five times counts five, so a page can post a high CD by repeating a small vocabulary rather than covering the subject, which is the exact behavior the metric is supposed to catch. So we measure a second number that counts each term once however often it recurs.

Unique Contextual Density (UCD) = (U(KV) + U(RE) + U(LSI)) / (TW − SW) × 100

Only the numerator changes. U(KV) is the count of distinct keyword variations present, U(RE) the count of distinct entities, U(LSI) the count of distinct co-occurring terms. Same denominator, same page, different question. CD asks how much topical work the text is doing. UCD asks how much of the subject the text has touched at all.

The repetition index, and what it exposes

Repetition Index = CD ÷ UCD

Take a page of 1,000 useful words where the subject’s terms appear 90 times. Its CD is 9%. If those 90 occurrences come from only 30 distinct terms, its UCD is 3%, and the repetition index is 3: each covered term recurs three times on average.

An index near 2 or 3 describes a page that keeps introducing new vocabulary as it goes. An index of 5 or 6 describes a page returning to a small set of terms over and over. Two pages can post an identical CD and have arrived there in opposite ways, and the repetition index is the only one of the three numbers that tells them apart.

In the table above, img.courses reaches the highest density in the set with an index of 4.87, while marketmuse reaches a lower density with an index of 3.07 on a page two-thirds the length. Those are different editorial strategies and neither is a verdict. A top-ranking page is by definition doing something right, and the index describes how a page is built rather than whether it deserves to rank.

UCD has a weakness of its own, symmetrical to CD’s. Its denominator is length, so it favors short pages. A 400-word page placing 17 distinct terms will show 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 number alone is a complete reading.

Breadth times density, not breadth or density

Breadth is how much of the subject’s vocabulary the page uses at all, which is what UCD’s numerator counts. Density is how much of the page is made of that vocabulary. The relationship between them is multiplicative rather than additive, because a page scoring near zero on either one is not rescued by the other.

Broad but not dense is a page mentioning everything once inside forty thousand words of filler. Dense but not broad is a page repeating six terms until the percentage looks respectable. The product is low in both cases, and this stays an approximation: it describes what relevant pages tend to look like when measured, not a metric Google has told us it computes.

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 with filler and generic high-quality content language drives the percentage down. Keyword stuffing drives it up while breadth stays flat, which is why the repetition index belongs on the report next to it.

What eighteen months of correlation data says about relevance

We ran the same correlation analysis in March 2025 and again in August 2026 to see whether the pattern was a snapshot or a trend. The data comes from Cora, from published correlation studies, and from our own measurement scripts run against live competitor pages. Negative coefficients indicate a link with better positions.

The factors that moved between 2025 and 2026

Three of the top four factors measure the breadth of vocabulary a page uses. The fourth measures how much of that vocabulary it shares with the pages already ranking, and it is now the strongest factor in the table.

One decline is worth naming rather than smoothing over. Entities in the title tag fell from −0.25 to −0.18, the largest movement of any factor in both exports. Whole-page entity coverage leads instead, and title placement now sits near the bottom of the on-page factors. Titles did not stop mattering. Stacking every entity you can fit into one was always the wrong lesson from the 2025 number.

What word count is doing in that table

Word count appears at −0.21 and we report it because it is in the export. We do not accept it as a ranking factor.

Covering a subject requires a certain number of contextual terms, entities and keyword variations, and placing those terms in readable sentences takes words. Content length is therefore the residue of semantic coverage rather than the reason a page ranks, which explains why the correlation is real but middling and why padding a thin page to a word target does nothing: the words arrive without the terms. Nobody can prove causation from correlation data, and between “write more words and Google rewards you” and “cover the subject and the words follow”, the second is consistent with every other factor in the table.

The honest limits of a correlation table

Cora’s dataset is built from the queries its own users run, which skews heavily toward the English-language US web. It is not a census of the internet, and no correlation tool has one. Search volume shifts continuously, so the composition of any top-100 sample moves underneath the measurement.

Google also changes its ranking systems constantly. In the last year for which it published complete figures it reported 3,234 launches out of 654,680 experiments, roughly nine changes to Search a day. Against that, a coefficient moving a few hundredths between two pulls eighteen months apart is the expected result.

What survived both readings is the ordering of the categories rather than the precision of the numbers. Semantic variation, entity coverage and contextual density sit at or near the top in both. Raw backlink count sits far down in both.

Topical relevance in AI search

Answer engines changed which half of this problem matters, and most content strategy written before 2025 has not caught up. Generative engine optimization is the name the industry settled on, and the mechanics below are why subject relevance carries more weight there than in classic organic search.

Query fan-out splits one prompt into many, and rewards vocabulary

An answer engine given a single prompt does not run a single search. It decomposes the prompt into multiple sub-queries, retrieves for each, and synthesizes across the results. Google has described this behavior for AI Overviews, and the pattern is visible in ChatGPT’s and Perplexity’s search traces.

The consequence for relevance is direct. A page that covers a subject with one phrasing can be retrieved by one sub-query. A page that covers the same subject with twenty-four distinct phrasings, the entities the subject involves, and the vocabulary that surrounds it, is eligible for retrieval by many of them. Fan-out multiplies the value of breadth in a way that classic ranking never did, because classic ranking gave you one shot per query.

This is the same finding as the correlation table, arriving from a different direction. Distinct variations, distinct entities and distinct co-occurring terms are what get a page retrieved. Repetition is not.

Why an answer engine cites a dense page and skips a broad thin one

Retrieval gets a page considered. Citation requires something further: the passage the model pulls has to answer the sub-query on its own, without the surrounding page.

A page with high breadth and low density mentions each term once in passing, so every passage containing a term is a passage that does not explain it. A page with both covers the term and says something specific about it in the same three sentences. The second is what gets quoted, because it is what can be quoted.

We have watched this on our own properties and on client sites, and we track AI visibility on our accounts rather than inferring it. The pattern is consistent enough to act on and we have not isolated it from other changes on those pages, so we hold it as an operational observation and not a measured result.

Brand recognition is the part on-page relevance cannot supply

The 2024 documentation exposed a site-level attribute named siteAuthority applied in Google’s Q* ranking system, and a chromeInTotal attribute recording site-level Chrome views. Rand Fishkin’s first recommendation on reading those documents was to build a recognizable brand outside Google search.

The outside evidence points the same way. An Ahrefs study of 75,000 brands published in May 2025 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 relevance gets you retrieved and brand gets you chosen, and no amount of vocabulary work substitutes for the second. We would rather state that boundary than sell page optimization as the whole answer.

Building topical relevance, in the order that works

Start from the SERP’s vocabulary, not your own

The strongest factor in the 2026 correlation data is the number of factors a page shares with the pages already ranking. That is not an abstraction. It means the benchmark for your page is the vocabulary the top results already use, and it is extractable.

Take the top five comparable pages for your target keyword. Extract every keyword variation, entity and co-occurring term they use, merge into one list, and count how many of the five use each term. Terms all five use are obligations. Terms one or two use are opportunities. Terms none of them use, that the subject genuinely requires, are where a page can separate itself, which is the whole reason this page talks about embeddings and nobody else’s does.

Keyword research comes first, and it is doing a different job here than usual. You are not hunting for a keyword to target. You already have one. You are mapping the subject around it. Start from the seed keyword, pull the phrase-match and question reports in Semrush or whichever tool you pay for, and read the long-tail keywords for the sub-questions they expose rather than for their individual volume. Google Trends will tell you whether interest in a phrasing is rising or dying. Google Search Console will tell you which search queries your existing pages already surface for, which is usually a wider set of relevant keywords than anyone expects and the fastest content gap analysis available to you. Moz, Ahrefs and Semrush disagree with each other on almost every number they publish, so treat any single tool’s figure as an estimate and the direction as the finding.

Then read the search engine results pages themselves. The ten results are Google’s own answer to what the subject contains and what user intent each search query serves, and they will contradict your keyword tool more often than you would like. Filter the list against your target audience before you write: a relevant topic that no prospective client would ever read is a relevant topic you should skip.

The right number of terms is set by the market, not by a rule. A competitive commercial query may demand a list of 180 terms. A narrow local one may demand 60. There is no threshold we have ever found, and we have looked.

Topic clusters are the site-level half of the same problem

A single relevant page sits alone. A cluster of them tells a search engine that the subject is what your site is for, which is the siteFocusScore half of the measurement.

Build one pillar page covering the subject broadly, and supporting pages covering its components in depth. That is the hub and spoke arrangement, and the pillar content is the hub. The pillar links down to each supporting page, each supporting page links back up, and related supporting pages link across where the connection is real rather than manufactured. A topic cluster built this way gives comprehensive coverage of a subject matter area that no single page could hold, and it is what site-level topic coverage looks like in practice.

Two failure modes are common enough to name. The first is a cluster of pages that all cover the same subtopics in slightly different words, which produces cannibalization rather than coverage: two of your own URLs competing for one query, and Google choosing whichever it prefers. The second is a pillar that links down to spokes which never link back, leaving the supporting content orphaned from the structure it is supposed to reinforce.

Where a variation deserves its own page rather than a section is a judgment about search intent, not about string similarity, and our page on keyword variations works through where that line sits.

Internal links are how relevance travels between pages

Variations in link anchor text correlate at −0.22, higher than word count. Anchor text is one of the few places where you state, in your own words, what a page is about, and generic anchors waste it entirely.

Write anchors that name the subject of the destination page. Keep the click depth of important pages to three or fewer from the homepage. Audit for orphaned pages, because a page nothing links to is a page your site is not vouching for.

Internal linking is also how crawlers find things. Googlebot follows links, and a page reachable only from a sitemap is a page competing for crawl budget it may not get. URL structure, headings, header tags and the meta description are all places where you restate the subject in your own words, and schema markup does the same job in a form built for machines: structured data naming the entity a page is about removes ambiguity that prose leaves open. None of these outrank vocabulary. All of them are cheap.

External links carry a version of the same signal. Link building on topically unrelated sites buys you very little now, and a smaller number of links from pages that genuinely cover your subject is worth more than a larger number from pages that do not. Our working hypothesis, which we have not isolated in correlation data, is that the strongest links come from pages with high contextual density of their own, because a semantically rich page endorsing you is a more credible endorsement than a thin one.

Off-topic content dilutes, it does not merely fail to help

This is the part most content teams get backwards, and siteRadius is why.

If a search engine measures how far each page sits from a site’s central theme, then publishing a page outside that theme does not just fail to earn anything. It moves the average, and pages far from the center dilute what the site as a whole reads as being about. A roofing company ranking well for “best rooftop bars in the city” has a traffic number and a relevance problem, and the traffic is the reason nobody fixes it.

The test we use is the one that keeps clients honest: would the person reading this page ever hire you, or would your own team send it to a prospect? If neither is true, the page belongs somewhere else. Auditing an existing site against that question usually means redirecting or rewriting a substantial share of the blog, and it is the single highest-return piece of work on most of the sites we inherit.

A content audit for this runs on data you already have. Pull your most-viewed pages from Google Analytics over the last twelve months and put organic traffic next to conversions. Pages with high traffic and no conversions are the candidates. Then decide per page: refresh it toward the subject you actually sell, redirect it to a page that belongs, or remove it. Time on page and user engagement are worth a glance, and they are weak evidence compared with whether anyone reading it ever became a customer. Thin content that is on-topic is usually worth rewriting. Well-written content that is off-topic is usually worth removing, and removing it is the decision teams resist hardest, because the traffic number feels like an asset.

There is a limit on this we should name. We do not know where the boundary of topical relevance sits. Mentioning Shakespeare in an article about content writing seems relevant to a person, and may register as noise to the algorithm. Some carefully chosen contextual terms are probably diluting rather than helping, and we cannot currently tell which.

How to measure a page you already have

Nothing in this method requires a subscription. It requires counting, and counting is arithmetic.

1. Pull the top five pages ranking for your target keyword that are the same page type as yours. Exclude directories, listicles and user-generated results, which rank on platform rather than text.
2. Extract every keyword variation, entity and co-occurring term from those five pages. Merge into one list. Record how many of the five use each term.
3. Count the body words of each competitor page, excluding navigation, footer and markup. Subtract stop words.
4. Count how many times each term on the merged list appears in each page. Divide by the non-stop word count, multiply by 100. That is each competitor’s contextual density against a shared list.
5. Count how many distinct terms from the list each page uses at least once. That is its breadth. Divide density by unique density for the repetition index.
6. Run the identical count on your own page against the identical list. The comparison is only valid if the list and the method are the same for every page measured, including yours.

Your page should exceed the highest measured competitor on both density and breadth. Close a density gap by adding a term the page has not used, never by repeating one it already has, because repetition raises density, leaves breadth flat, and moves nothing the correlation data links to rank.

The published version of this method, with the full derivation and the pillar weights, is the SEO Contextual Density Index. The entity half of the term list is worked through in our page on entity SEO.

What topical relevance will not do for you

Contextual relevance work cannot compensate for a page a search engine struggles to crawl, index or render. Indexation, canonicalisation, Core Web Vitals, mobile rendering, HTTPS and click depth are prerequisites, and semantic work on a broken foundation returns nothing.

It cannot supply brand recognition, and the AI Overviews data above suggests brand is now the heavier signal for citation in generated answers.

It cannot substitute for topical expertise. E-E-A-T is loosely defined and it is at least Google’s own framework, and the trustworthiness half of it comes from who wrote the page, what they have done, and who says so elsewhere. Vocabulary coverage will not make an inexperienced author read as a practitioner. We also will not reason from vendor authority scores here, because domain authority and page authority are each vendor’s own invention, they disagree with each other, and at least one of them takes organic traffic as an input, which makes explaining a poor ranking with a low score circular.

It cannot survive contact with engagement signals. The leaked documents describe NavBoost, a system using click and dwell behavior, and no amount of vocabulary makes a page people leave immediately look good to it.

And it is a proxy. Everything on this page is a description of what relevant pages tend to look like when measured, validated against roughly 80% of the client implementations where we have applied it and the technical foundation was sound. That is an operational observation from our own client base, not a controlled study, and we state it as such.

Frequently asked questions

What is topical relevance in one sentence?

Topical relevance is how closely a page’s vocabulary matches the vocabulary a search engine associates with the subject the page claims to cover.

Is topical relevance the same as topical authority?

No. Topical relevance is a measurable property of a single page, established by the terms it uses. Topical authority is a site-level reputation built over time from coverage, citations, links and brand searches. A relevant page on a new site will rank for specific phrasings and stall on competitive head terms, because authority is what carries the head terms.

Is topical relevance a Google ranking factor?

Google has never published a factor by that name. What the evidence supports is narrower: the 2024 documentation leak names attributes measuring topical concentration, Google’s retrieval systems demonstrably operate on semantic representations rather than string matching, and the breadth of a page’s subject vocabulary correlates with position more strongly than any other on-page property we measure. That is a strong prior and it is not proof.

How is topical relevance different from keyword density?

Keyword density counts occurrences of one string against total words. Topical relevance concerns the whole vocabulary of a subject, of which the exact keyword is a small fraction. The exact term still has to be present in the title, the H1 and the body, and its frequency beyond that correlates weakly. Distinct variations correlate more strongly than repetition of the core term.

How many words does a topically relevant page need?

There is no word target. The number of contextual terms the subject requires is set by what the ranking pages already cover, and the word count is what it takes to place those terms in readable sentences. Competitive commercial queries can demand 180 terms and narrow local ones 60, and the resulting lengths differ accordingly.

Does off-topic content really hurt my other pages?

The leaked siteRadius attribute describes a measurement of how far each page sits from a site’s central theme, which implies an average that off-topic pages move. We have not isolated the effect experimentally and nobody outside Google has. Our position is that content earning traffic from people who will never become customers is worth removing on business grounds regardless of whether the algorithmic dilution is real.

Does topical relevance matter for ChatGPT and AI Overviews?

It matters more there than in classic ranking. Answer engines decompose a prompt into several sub-queries and retrieve for each, so a page covering its subject with many distinct phrasings is eligible for far more retrievals than one covering it with a single phrasing. Getting cited then requires passages that answer a sub-query on their own, which needs density as well as breadth.

How do I measure topical relevance without buying a tool?

Extract the terms from the five top-ranking comparable pages, merge them into one list, and count occurrences of that list in each page and in yours, using the same list and the same method throughout. The arithmetic is in the section above. A spreadsheet is enough.

Measure your pages before you rewrite them

Most pages that underperform are not badly written. They are narrow, and nobody has counted how narrow. ProStar SEO measures contextual density and breadth on every page we touch, against the pages currently ranking, before a word of the content marketing gets changed.

If you want to know where your pages sit against the market, get in touch and we will run the measurement on three of them.