How do I organise our content so AI assistants see us as the expert?

You have forty or fifty articles on the website, some good, some written by an agency years ago and never read since. When a buyer asks ChatGPT or Gemini who the best specialists in your field are, a competitor with half your experience gets named and you do not.

In our experience this is rarely a volume problem. It is an organisation problem. If forty pieces point in forty directions, disagree with each other in places and answer half of each question, an assistant has no single page or passage it can rely on to say what your firm knows.

There is also a mechanical reason organisation matters more than it did. Google's AI features split one question into several related searches, a technique Google calls query fan-out, and cite pages that answer the parts. Ahrefs found in March 2026 that only 38% of AI Overview citations came from pages ranking in Google's top ten for the original search, down from 76% in July 2025. The rest ranked lower or not at all for it, a shift that fits Google running several related searches for each question. A group of pages that covers a subject and its follow-up questions gives an assistant more to cite.

This guide explains how to arrange what you publish into topic clusters that buyers and assistants can follow, using a worked example from employment law and the changes we made to our own guides.

#What is a topic cluster?

A topic cluster is a main guide on one subject, linked to a handful of articles that each cover part of that subject in more depth. The main guide is usually called the pillar page; the others are child pages. The pillar links to every child, every child links back, and children link to each other where it makes sense.

The idea was popularised by HubSpot and built for Google, which reads the link structure as a sign that you cover a subject in depth. AI assistants need a stricter version, because they quote passages rather than weighing links alone.

#What is query fan-out, and how does it change content planning?

Query fan-out is Google's term for its AI model issuing "a set of concurrent, related queries" to gather more material for one question. Google's guide to its AI features, updated in July 2026, gives the example of "how to fix a lawn that's full of weeds" fanning out to "best herbicides for lawns", "remove weeds without chemicals" and "how to prevent weeds in lawn".

Google says AI Mode, available in the UK since July 2025, uses this technique. For a B2B firm, a buyer's single question, "how do we handle redundancies without ending up at tribunal?", is answered from pages on consultation, selection, pay and settlement agreements. A clear page for each gives you several chances to be cited; one long page that touches each briefly gives you one.

There is a limit. The same Google guide warns that creating "separate content for every possible variation of how people might search", including fan-out queries, "primarily to manipulate rankings or generative AI responses" breaks its scaled content abuse spam policy. A cluster answers the real sub-questions of a subject once each. It is not a page for every phrasing.

#How do AI assistants read a cluster?

Mostly one passage at a time. When an assistant answers a question it retrieves the passages that best match each part of it, often from different pages and different sites, and composes an answer. It does not read your cluster from start to finish.

So each page, and ideally each section, should open with a direct answer that makes sense without the rest: a 2026 analysis of 1.2 million ChatGPT answers by Kevin Indig found 44% of citations came from the first 30% of a page. The pages must agree, because an assistant that finds two versions of the same fact on your site has a reason to use neither; our guide to why ChatGPT describes firms wrongly covers what happens then. And links should say what they mean, for readers and crawlers alike.

Language models handle synonyms well, so "buying committee" on one page and "stakeholders" on another are unlikely to confuse them. Consistent terms still matter for a simpler reason: a reader and an assistant can quote your framework accurately only if you describe it the same way everywhere.

#What does a cluster that assistants can use look like?

Every page is on the same subject, the children develop the framework the pillar sets out, and every page could be quoted on its own. Here is the example we use.

A firm of employment lawyers wants to be known for handling redundancies. The pillar sets out a four-stage process; each child covers part of it.

• Pillar: How to run a redundancy process that stands up at tribunal (sets out the four stages)
  • Child 1: Collective consultation: when it applies and how to run it (stage one)
  • Child 2: Selection criteria and scoring: what is fair and what is not (stage two)
  • Child 3: Redundancy pay and settlement agreements (stage three)
  • Child 4: Redundancy in a business acquisition (a situation that changes every stage)

A loose cluster would add "how to write an employee handbook" or "managing poor performance". Those are employment law topics, but they say the firm writes about employment law in general, not that it is the firm to call when redundancies are coming.

The children should develop the pillar's framework, not start their own. If the pillar says there are four stages, the best children are one deep article per stage. "Ten redundancy tips", "a redundancy case study" and "redundancy software", each with its own structure, read as three unrelated approaches rather than one considered method.

#How should the pages link to each other?

Use three kinds of link, and make every anchor say what the linked page is. Each type does a different job.

Link mid-sentence where the idea comes up, link in both directions, and only link where the connection means something. A block of "related reading" at the bottom tells nobody why the pages belong together, and padding a page with links makes the cluster look less focused. How the rest of your site is organised matters too; our guide to organising a website so buyers and AI understand it covers navigation and page hierarchy.

#How do I find the follow-up questions buyers ask about my service?

Start with the questions buyers already ask you, then check what the assistants and search engines add. Each follow-up that has a real answer becomes a child page or a section of one.

#Step 1: Mine your own conversations

List the questions from sales calls, enquiry emails and proposals. The second and third questions a prospect asks are usually the follow-ups.

#Step 2: Check Google's suggestions

Type your main question into Google and note the autocomplete suggestions and the "People also ask" questions. They show the related searches people run.

#Step 3: Ask the assistants what comes next

Put the buyer's question into ChatGPT, Gemini and Google AI Mode, note which sub-topics the answers cover, and ask "what would a buyer need to know next?".

#Step 4: Map each question to one page

Give each real follow-up question one home: a child page if it needs depth, a section of the pillar if not. Merge any two pages that answer the same question. Our guide to writing content around the questions customers actually ask covers the research in more detail.

#How should the pillar and the children be written?

The pillar defines the subject, names the three to five parts of the framework, gives a short section on each that points to the child going deeper, and covers common mistakes and how to tell it is working. It does not go deep on everything.

Each child opens by placing itself in the framework ("stage two of the four-stage process"), answers its question in the first two sentences, goes deep and links to its neighbours. Comparisons, step-by-step guides and worked examples make good children; label any illustrative example as such. Keep a short glossary of the cluster's terms, with one definition each, so every writer uses the same words; our guide to building a glossary AI assistants use as a reference explains how.

#Worked example: what we changed in our own guides

We applied the same test to our own Insights section. The topic on being named by AI assistants has a topic page acting as the pillar and a dozen guides as children. Read in one sitting, they disagreed on points a buyer cares about: one guide said a B2B firm should start with Perplexity and Claude while another said to check every assistant, and guides gave different answers on whether Google rankings carry over and how many questions to track. We settled one answer for each and rewrote every guide to give it.

Each child now does one job: how to get ChatGPT to recommend your firm is the foundation, SEO versus AI search the comparison, and how AI crawlers read your site a technical piece.

#What does this look like on a larger site?

The same principles hold at scale: a structure that follows how a buyer thinks, and facts that agree across every page. When we rebuilt the website for arum, which advises banks, governments and major brands on collections technology, the result was "310 pages rebuilt across three regions, in weeks", in "3 regional versions: UK and Europe, North America, Australia and New Zealand."

The structure followed the buyer: "Services organised the way a client buys them: five stages from advice to optimisation, plus three specialist services." And "Each of the 12 case studies leads with a number, and each of the nine sector pages names the regulation in play." Underneath sat "Proper structure and markup on every page, a plain summary AI assistants can read, facts stored once so two pages never disagree, and every old address redirected so years of rankings carry over." The arum case study has the detail.

#How do you tell whether a cluster is working?

Run a fixed list of 20 to 50 buyer questions on the subject, each several times a month in a clean session, and record who is named and which of your pages are cited. Answers vary from run to run, so count a win only when it repeats. Our guide to finding out whether ChatGPT recommends your firm sets out the method.

Watch four signs: how often you are named against your closest competitors; whether several cluster pages are cited, not always the same one; answers that repeat your framework back ("there are four stages to a fair redundancy process"); and whether enquiries that start from cluster pages are better qualified.

On timing, assistants that search the web live (ChatGPT search, Perplexity, Copilot and Google's AI features) can reflect new or rebuilt pages within weeks. Answers drawn from a model's own memory change only when the model is retrained, which takes months.

#What to do this month

1. Choose one subject you want to be the first name for: one where you have real depth and the work is profitable.
2. List every page you have on it and decide which is the pillar, which are children, and which should be merged, moved or retired.
3. Write down the framework the pillar will set out, and a one-page glossary of the terms every page will use.
4. Collect the follow-up questions from sales calls, Google's suggestions and the assistants, and give each one a single home.
5. Read the cluster in one sitting and fix every contradiction between pages before you write anything new.
6. Record a baseline with 20 of your buyer questions in ChatGPT, Google AI Mode and Copilot, and repeat monthly.

#Frequently asked questions

#How many articles do I need in a cluster?

Start with a pillar and four or five children. That covers the main follow-up questions and stays easy to keep consistent. If a subject needs ten or more, split it into two clusters.

#Should each of our service pages be a pillar page?

Usually not. A service page explains what you sell and asks for an enquiry; a pillar explains the subject to someone still researching. Link them both ways. A firm with one core service can use its service page as the hub, provided it answers the subject and does not only sell.

#Do topic clusters still work for Google in 2026?

Yes. Google says SEO best practices "continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems", and its AI features run several related searches per question, which suits sites that cover the follow-ups. What does not work is a page for every phrasing, which Google treats as scaled content abuse.

#Should our guides sit on a blog or in a separate resources hub?

A hub organised by topic, with one page per topic acting as its pillar, is easier for readers and crawlers to follow than a date-ordered blog. The folder name in the address matters far less than the links between topic pages and guides.

#What if my subject does not break down into a pillar and children?

Then it may not need a cluster. A narrow subject is often better served by one thorough article, and forcing it into five short pages that repeat each other is worse. Our guide to why ChatGPT ignores sites full of thin pages explains why.

#How often should I update cluster articles?

Review each quarter: refresh examples and figures, and date each real change. Seer Interactive found in July 2026 that 75% of pages cited by AI assistants had been updated in the last 12 months, in a study of four consumer and finance industries. Once a year, read the whole cluster together and check the framework still holds.

A well-organised cluster is one of the most reliable ways we know to move a firm from sometimes mentioned to named first on a subject. To see how your content reads to buyers and AI assistants today, book a free Growth Review: 30 minutes with David Adams and three fixes in writing. Rebuilding a site around the questions buyers ask is the work of Website That Sells.

#Sources