Topic Clusters for SEO: Google Reads the Links, Not the Diagram

What is a topic cluster in SEO?

Google publishes a guide to the ranking systems it uses, from RankBrain to the site diversity system that limits how many results one site gets. Topic clusters are not on the list, and neither is topical authority.

A topic cluster is a group of pages on one subject: an overview page, usually called the pillar, that answers the broad question; supporting pages that each answer one narrower question in depth; and the internal links that join them. It is a way of organizing a site, with no score of its own. What it changes is two things Google does document: how Google finds your pages and reads the relationships between them, through links and the words in them, and whether your site answers the questions a searcher on that subject actually has.

HubSpot, which popularized the idea, still describes the model as related cluster pages organized around one central pillar page, with internal links connecting every cluster page back to the pillar. The hub-and-spoke diagram that usually comes with it is the least useful part. The decisions behind it are smaller, and each can be checked: which searches share a page, which page owns each search, which sentence links where and with which words, and whether traffic moved compared with pages nobody touched.

Do topic clusters help SEO? What Google says links do

On links, Google is unusually specific. Its link best practices say Google "uses links as a signal when determining the relevancy of pages and to find new pages to crawl," and that "every page you care about should have a link from at least one other page on your site." Good anchor text, the same page says, is "descriptive, reasonably concise, and relevant to the page that it's on and to the page it links to," and the words before and after a link matter too.

Discovery is the plainest effect. The SEO Starter Guide says Google "primarily finds pages through links from other pages it already crawled." A supporting page that nothing links to may still be found through a sitemap, but without any context from the pages around it.

Relationships come next. Google's documentation on ecommerce site structure says it "tries to find the best content on your site by analyzing the relationship between pages based on their linkages," and uses links "to infer the relative importance of a page over the rest of your site." The same page settles a common cluster argument: "Google generally doesn't look at the structure of URLs to work out the structure of a site." Putting supporting pages under a /severance/ folder does not make them a cluster in Google's eyes; the links do. (The Starter Guide suggests folders on sites with more than a few thousand URLs, but to help Google learn how often each section changes.)

Coverage is the second effect, and here the guidance is editorial. Google's page on helpful, people-first content asks whether a page offers "a substantial, complete, or comprehensive description of the topic" and whether a site has "a primary purpose or focus," and it flags producing lots of content on many different topics in the hope that some of it performs. On whether Google has a preferred word count, it answers: "No, we don't."

"Topical authority" sells most cluster advice, so it is worth being exact. The only system Google has named topic authority, described in a May 2023 post, helps surface expert sources for news queries in areas such as health, politics and finance. Google has not documented a general topical authority score for websites; scores under that name in SEO tools are the vendors' own metrics. HubSpot traces its model to its content team's finding that pages climbed as they added internal links between related pages. That fits what Google says, and it is one company watching its own pages, which is why measuring a cluster matters as much as building it.

How keyword clustering works: SERP overlap versus word similarity

Before a cluster has pages it has a keyword list, which has to be cut into groups that each become a page or a section. There are two ways to cut it, and they fail in different places.

Clustering by SERP overlap

For each query, collect the top 10 organic URLs Google returns in the market you care about. For each pair of queries, count the URLs that appear in both lists. If the count reaches a threshold, the queries go on one page. The reasoning: Google has already decided which documents satisfy each search, and when the same documents satisfy both, it is treating them as one job. Keyword Insights, one tool that clusters this way, documents a default of three shared URLs; its worked example groups "weight loss with shots" with "weight loss with injections" because their top 10 results share four.

The threshold is a trade-off, and each setting has a predictable failure.

The comparison method matters as much as the number. Keyword Insights describes two: a centroid method, which compares every keyword with the group's highest-volume keyword and "generally results in larger clusters," and an agglomerative method, which compares all keywords against one another for "smaller, tighter clusters." In either output, look for chains, where A shares four URLs with B, B shares four with C, and A and C share none. Whether C lands with A depends on the method; the reliable fix is to open the three results pages and look.

Two other settings change the answer. Count URLs, not domains: a large site that ranks one of its pages for the first query and a different page for the second has already made the split you are deciding. And check the right market. A Toronto law firm needs Toronto results.

SERP overlap has blind spots. It sees a snapshot, so volatile results make volatile clusters. It needs results to compare, and the long tail is where data runs out: queries Search Console anonymizes, which Google defines as those not issued by more than a few dozen users over a two-to-three month period, never appear in an export at all.

Clustering by word similarity

Word-similarity clustering groups queries that share words or, in newer tools, sit close together in a language model's representation of meaning. It is fast and works on any list, including queries no tool has results for. It fails where mistakes cost money. "Severance package Ontario" and "severance package lawyer Toronto" share most of their words, yet one searcher wants to understand an offer and the other wants to hire someone. Meaning-based models handle synonyms better, but they still measure what words mean, not what Google ranks.

So use the two in sequence: word similarity to sort a long list into rough piles, then SERP overlap on the pairs where the answer decides one page or two. Our guide to keyword research with AI covers the first pass with Claude.

One page or two: how clustering decides cannibalization

Keyword cannibalization is two of your own pages competing for one search so that neither does well. Clustering gives you the rule before anything is written: queries that share enough URLs get one page, and queries that do not get separate pages. A second page for a query that belongs to the first creates the problem on purpose.

For pages that already exist, Search Console shows what Google has done. Filter the Performance report to the query and open the Pages tab. Two of your pages earning impressions is not a problem in itself: Google's ranking systems guide says its site diversity system "generally won't show more than two web page listings from the same site in our top results," and two strong listings is a good outcome. The pattern to fix is two pages trading places week to week while neither holds a good position.

The fixes differ in how much they change. The lightest is to re-scope one page to a different job, with a title and headings to match. The heaviest is to merge the weaker page into the stronger and redirect it permanently; Google's redirect documentation says its indexing pipeline uses a permanent redirect "as a signal that the redirect target should be canonical." Either way, point the site's internal links at the owner. If half your links about severance calculations go to a guide and half to an old blog post, the site gives Google two answers about which page explains the subject. That last point is an inference from Google calling links a relevancy signal, not something Google has spelled out.

A topic cluster example: Harbour & Finch (sample client)

Harbour & Finch LLP is a sample client, a Toronto employment law firm used for demonstrations across this site; every page, query and number in this section is sample data. The subject is severance. The firm has a service page offering severance reviews, a 2023 blog post on how severance is calculated, a newer guide covering the same calculation, and nothing on negotiating an offer or on Employment Insurance.

An overlap check on Toronto results, at three shared URLs, sorts the list like this.

The calculation group is the cannibalization case: two existing pages, one job. The newer guide is stronger in this sample, so it keeps the job and takes anything useful from the old post, and the old URL redirects to it permanently. The lawyer queries share only two URLs with the negotiating group and show a local pack, so they stay with the service page.

What goes on the pillar, the overview page

The pillar answers the broad question well enough that a reader who wants only the outline can stop there, and it routes everyone else. For Harbour & Finch that means what severance is, the difference between the statutory minimum and what a court may award at common law, what to do before signing anything, and a route to the review service. Each narrower question gets two or three sentences and a link to the page that answers it in full. A pillar that only lists links is a table of contents with nothing to read; one that answers everything in full leaves its supporting pages nothing to add.

What goes on a supporting page

A supporting page goes deep on one job: the worked calculation with the inputs that change the answer, the edge cases, the documents to have at hand, the questions only someone at that stage asks. A question earns its own page when the overlap check says Google treats it as a separate job and the full answer would bury the pillar for readers who do not need it. If two paragraphs on the pillar would satisfy that searcher, it is a section.

The links, and the words in them

The anchors name the destination in words a reader would use, as Google's guidance asks, and they differ because the sentences around them differ, not to satisfy a rule. Every page gets at least one link from inside the cluster, in the body where the next question comes up rather than only in a footer. And the service page belongs in the cluster, because the reader the cluster serves eventually needs it.

To check what a page already links to, paste its HTML into our free internal link analyzer. It lists every internal link with its anchor text and flags empty or generic anchors and nofollow links, so you can confirm the pillar links to each supporting page, and with which words.

A topic cluster template you can copy

Fill in one block per page, top to bottom. The order is deliberate: the queries come before the title, and the reason a page exists comes before what it covers.

TOPIC CLUSTER
Subject (in the searcher's words):
Market checked (country, city, device):
Overlap threshold (shared URLs in the top 10):
Owner:                          Live date (YYYY-MM-DD):
Search Console page regex:

PILLAR (overview page)
URL:
Primary query:
Other queries it owns (overlap at or above threshold):
Searcher's job, in one sentence:
Short answer to each supporting question (2 or 3 sentences each):
Links out (anchor text -> URL, and the sentence it sits in):
Link to the service or product page:
Status: exists / rewrite / new

SUPPORTING PAGE (copy this block for each one)
URL:
Primary query:
Other queries it owns:
Searcher's job, in one sentence:
Why this is not a section of the pillar:
Must cover:
Link to the pillar (anchor text, sentence it sits in):
Links to sibling pages (anchor text -> URL):
Pages merged into it (old URL -> this URL, permanent redirect):
Status: exists / rewrite / merge / new

CHECKS BEFORE THE LIVE DATE
[ ] No two pages in the cluster share a primary query
[ ] Every page has an in-body link from another page in the cluster
[ ] Every redirect is permanent and points at the owning page
[ ] Links elsewhere on the site about this subject point at the owner
[ ] Every page is indexed (URL Inspection in Search Console)

MEASUREMENT
Before window:                  After window (same length):
Control pages (untouched, similar traffic, no new links in):
Cluster clicks, before / after:
Control clicks, before / after:
Cluster impressions, before / after:
Control impressions, before / after:

Two fields do most of the work. "Why this is not a section of the pillar" stops the cluster from becoming a page for every phrasing; if the honest answer is "more keywords", it is a section. A page per phrasing is also the pattern Google's spam policies call scaled content abuse when the pages add little: "many pages are generated for the primary purpose of manipulating search rankings and not helping users." The other field is the live date, which makes the cluster measurable.

How to tell if a topic cluster strategy is working

A cluster is a change to a set of pages on a known date, so it can be measured like any other change. The trap is the simple before and after: compare the eight weeks after launch with the eight weeks before and you have measured the season, any core update in the window and the rest of the site's movement, along with the cluster.

1. In the Performance report for Search results, add a filter, choose Pages, then Custom (regex), and enter a pattern that matches only the cluster: /severance/ for a folder, or a list for scattered pages, such as /(severance-pay-ontario|how-severance-is-calculated|negotiate-severance-package|severance-and-ei)/?$. Google's help page on advanced filtering says the syntax is RE2, patterns match anywhere in the URL unless anchored with ^ or $, and matching is not case-sensitive by default.
2. Choose Compare and set two date ranges of equal length, one ending before the change and one starting after it. Skip the days it was rolling out, and use whole weeks so weekdays line up.
3. Do the same for a control: pages you did not touch, with similar traffic beforehand and none of the new links pointing at them. A page that gained a link from the new pillar has been changed, even if nobody edited it.
4. Compare the two changes. If the cluster's clicks rose 30% and the untouched pages rose 25%, most of the rise was the tide.

Filter by page, not by query. Google's post on Performance data says anonymized queries count in chart totals "unless you filter by query," and long-tail clusters live on those searches, so a query filter can hide much of a cluster's traffic. Read impressions next to clicks. More impressions with no more clicks means the pages are being tested for more searches and losing, which points at titles and answers rather than structure. And for each main query, the Pages tab should show the owner you chose collecting the impressions.

Two limits apply. Google reports AI Overviews and AI Mode traffic inside the Web search type, according to its AI features documentation, so a comparison cannot separate them. And small clusters cannot prove small effects, because a few dozen clicks a month can swing by a large percentage on noise alone. Our guide to proving an SEO change worked with a control group has the full method, including how to tell an effect from ordinary churn, and the free Did it work? tool runs it on two Search Console exports in your browser.

Topic clusters and AI search: what is known and what is guesswork

Searches such as "topic cluster optimization for AI search" assume clusters need a new version for AI Overviews, AI Mode and ChatGPT search, the work many teams now call AEO. What is documented is ordinary.

Google's documentation on AI features says "the best practices for SEO remain relevant" for AI Overviews and AI Mode and that "there are no additional requirements" to appear in them. A supporting link must be "indexed and eligible to be shown in Google Search with a snippet." Among the practices that still apply it lists "making your content easily findable through internal links on your website," and it says "you don't need to create new machine readable files, AI text files, or markup to appear in these features." The work that helps a cluster rank in ordinary results is the work Google says applies here.

The same page describes query fan-out: both features may issue "multiple related searches across subtopics and data sources" while building a response, which lets Google show "a wider and more diverse set of helpful links." For ChatGPT search, OpenAI's crawler documentation says OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features," and that sites opted out of it "will not be shown in ChatGPT search answers, though can still appear as navigational links." That documentation covers access, not how sources are chosen.

Past that, it is inference. If an AI answer is assembled from several subtopic searches, a site with one strong page per subtopic plausibly has more chances to be a source than a site with one thin page covering everything. That is a fair reading of how Google describes fan-out, and nobody outside Google has shown it holds for clusters. Claims that a cluster lifts "AI visibility" by a stated percentage, or that language models reward topical authority, have no support in anything Google or OpenAI has published. Our explainer on how Google's AI answers choose their sources goes through the evidence.

Because Search Console does not split AI traffic out, the direct way to know whether assistants name a client is to ask them the same questions repeatedly and count, with a margin of error, since answers change from run to run. AIO Copilot samples how often AI assistants name a client against its rivals and reports that margin next to the share.

How to build a topic cluster content strategy from the pages you have

Most sites already have the start of a cluster on every subject they cover: a few pages that grew up separately, linked by accident if at all. Pick the subject that matters most to the business, export its queries, run the overlap check on the pairs that decide one page or two, fill in the template, fix the links, and write down the date.

This page is a test of the same kind. In the 28 days before this rewrite, Google showed it about 2,080 times, at an average position of 24.5, for searches such as "topic cluster content strategy" and "topic cluster strategy". Nobody clicked. A page can match a subject closely enough to be shown for it and still give a searcher no reason to choose it over the results above. Whether this version does better is a question for the same method: its clicks after 22 October against posts nobody touched.

Frequently Asked Questions