How to Build Your First Citation Cluster, Step by Step (What I Got Wrong on Mine)

A citation cluster is a group of pages on your own site that each answer one buyer question about one narrow topic and link to each other, so that when someone asks ChatGPT, Claude or Perplexity a question inside that topic, your site is one of the sources it pulls the answer from. That is the whole thing. Here is how to build your first one, in the order I do it for clients now, plus what I did wrong on my own.

A quick note on where this comes from. The first cluster I ever built was for BakingSubs, a gluten-free baking site my girlfriend and I started from scratch with no audience and no ad budget. According to my Bing Webmaster data, Microsoft Copilot has recommended it more than 340,000 times and quoted more than 160 pages by name. That's my site and my subject, so I can't promise yours will look anything like that. But the first cluster on it was a mess, and the mess is the useful part.

Key takeaways

A citation cluster is one narrow topic plus a set of question-shaped pages that each answer a real buyer question directly and link back to each other. Narrow beats broad every time.
Pick the questions from the people who actually buy from you, not from what you assume they want. That was my mistake, and it cost me about a month and a half.
Every page opens with the answer, then shows the part that comes from actually doing the work. The AI needs something worth using in the answer.
Track whether the engines start naming you. My first cluster got zero Google traffic and still got cited. If I had only watched traffic, I would have called it a failure.

What happened on my first cluster

Here's the context. When I started BakingSubs, the biggest searches I could see were all substitutions: bread flour for all-purpose flour and back again, egg substitutions, sugar substitutions, dairy substitutions. So the first pillar was substituting flours, and the whole site was substitutions until we narrowed into dietary restrictions later.

The way we built it was to make every ingredient a database entry. About 100 ingredients, each with two or three substitutions, each one actually tested in a kitchen before it went up. Call it 300 pages. It took about a month and a half.

And those pages did not rank. They got absolutely cannibalized by Google's AI Overviews, because "can I substitute this for that" is exactly the kind of question Google now answers on the results page itself. Zero traffic. (Kill me.)

Here's what I didn't understand at the time. The first sign it was working had nothing to do with Google. People started finding the site anyway. I shared it on Reddit and got a bunch of replies saying the substitutions were accurate, that this was exactly what they would have used. A couple of people rated the substitutions on the site itself. And the first page an AI engine cited by name was the xanthan gum page, because xanthan gum shows up in almost every gluten-free substitution and the page said something specific about baking with it.

So a cluster that lost on traffic won on being the source. That is the distinction this whole method is built on, and I had to learn it by accident.

Step 1: pick one narrow topic you can actually out-answer

Choose a single narrow topic where you know more than the generic pages already ranking for it. The test I use now: write the topic as a sentence a buyer would say out loud. "How do I know if my roof needs replacing or just repairing" is a topic. "Roofing" is not.

The mistake I made was picking topics from search volume instead of from the person buying. I built what I thought people wanted. If I had built for the problems people were actually having, I think I would have gotten results a lot faster. For a roofer that means the specific repairs customers call about, what goes into replacing a metal roof with a shingle roof, the questions that come up right before somebody signs. Not "roofing tips."

Pick a topic where you have answers nobody else can copy: a number from your own work, a mistake you watch customers make, a judgment call you make differently from the rest of your field. That is the material an AI engine cannot summarize away, because it only exists on your page.

Step 2: map the questions buyers ask around it

List the exact questions a buyer asks before, during and right after choosing someone like you. Get them from where buyers actually ask, not from a keyword tool alone:

What people ask you in the first sales call, word for word
The objections that come up right before someone says yes or no
What people type into ChatGPT when they are comparing options
The "is X worth it", "X vs Y" and "how much does X cost" phrasings

Each question becomes one page. If two questions have the same answer, merge them. If one question needs a different answer for a different buyer, split it. One page, one question, one job.

This is the part that stalls most people, and it is the part I got wrong on my own site, so it is the part the AI Visibility Tracker does first: it finds the buyer questions AI is already being asked in your niche and shows you which ones name a competitor instead of you.

Step 3: write each page to answer the question first

Open every page with the answer, plainly, in the first sentence or two. Then expand. The engines pull that opening and repeat it, so a page that warms up for three paragraphs gets skipped for the one that answers in line one.

The structure I use on every cluster page now:

Two or three sentences that state the answer directly, no throat-clearing.
Three or four key takeaways as bullets, each a standalone fact an engine can lift.
The body, with each heading phrased as the question a buyer would actually type.
A short FAQ of three to five more questions, each answered in a few sentences.

Then the part that matters most. Put in something that only exists because you did the work. If your page just repeats the same stuff that's on nine other websites, there's not much reason for an engine to choose your page. But if your page contains information that comes from actually doing the work? That's different. On BakingSubs it was tested substitutions and one specific page about xanthan gum. For you it is your real prices, the mistakes you see customers make, the way you handle a weird situation, the answer you give every single time somebody calls. That stuff isn't sitting on 500 other websites.

Step 4: link the cluster together

Link every page to the pillar and to the two or three most related pages in the cluster, using a plain phrase from your own writing as the anchor. This is what turns a pile of posts into one structure an engine reads as authority on the topic. Skip it and you have written good pages that nothing connects.

Two rules keep it honest. Link only where the sentence would send a reader there anyway. And never stuff a keyword into an anchor. There is more on the wiring in how to build topical clusters AI engines actually cite.

Step 5: publish, then watch whether AI starts naming you

Publish the pages, then check directly whether ChatGPT, Claude and Perplexity begin naming you when someone asks a question inside your cluster. This is the step I skipped, and it is why I nearly wrote off a cluster that was working.

Ask the engines the questions your cluster answers, in the words a buyer would use, and note two things: whether you appear at all, and whether you are named or just blended into a generic answer. Do it on a schedule, because citations move week to week as the engines re-read your site. The free Visibility Check gives you the first read on where you stand today. The tracker watches it from there and tells you the day an engine starts saying your name.

And I'm not going to promise you that you publish a page on Tuesday and ChatGPT starts sending you customers on Friday. I don't know when an AI will recommend your business. Nobody honestly does. What I can tell you is that the first page cited on my site was not the one I expected, and I only found out because I went looking.

How many pages does a first cluster need?

Fewer than mine. Three hundred was a database, not a cluster, and most of those pages answered questions Google decided to answer itself. If I were starting today I would build the cluster around the questions buyers ask right before they choose, finish that set completely, and watch what gets cited before starting a second one. A finished cluster that gets you named is worth more than three half-built ones that get you nothing.

Frequently asked questions

Do I need to rank on Google first to get cited by AI?

No. My first cluster is the proof: those pages never ranked, Google's AI Overviews took the queries, and an engine still cited the xanthan gum page by name. AI engines pull from sources they judge useful on a specific question, which is not the same thing as your Google position.

How long until AI engines start citing a cluster?

I don't know, and I would not trust anyone who gives you a number. Mine took a while, and the first citation landed on a page I was not watching. Track the questions directly and on a schedule, and you will see it when it happens instead of guessing.

Can I build a citation cluster without any tools?

Yes. Mapping the buyer questions, writing one page per question with the answer first, and linking them together can all be done by hand. That is how I built mine. The AI Visibility Tracker exists because the two places I lost the most time, picking the wrong questions and not tracking the result, are the two things it does for you.

What is the difference between a citation cluster and a topic cluster?

A topic cluster is built to rank a pillar page on Google. A citation cluster is built so an AI engine has a reason to name you when it answers a buyer's question. The structure looks similar. The pages are written differently: answer first, and something on every page that came from actually doing the work.

If you have built one, or you are in the middle of one that isn't showing up anywhere yet, shoot me a message on LinkedIn and tell me what you are seeing. I read every one.