How to Build Topic Clusters That AI Engines Cite

Keyword-first content planning is breaking down, and there is a specific number that explains why. In an AirOps analysis of 15,000 ChatGPT prompts reported by Search Engine Land in March 2026, the engine expanded those prompts into 43,233 actual queries, and 95% of those fan-out queries had zero traditional search volume.

You cannot target queries that do not exist in a keyword tool. You can only cover the territory they come from. That is what a topic cluster is for.

Why clusters beat keyword lists now

Three findings from that same dataset make the case:

• 89.6% of prompts triggered two or more follow-up searches. One question becomes many.
• 32.9% of cited pages appeared only in fan-out results, never in the original query. A third of citations come from questions nobody typed.
• 55.8% of cited pages ranked in Google’s top 20. Classic ranking still gates eligibility.

Put together: you need conventional ranking strength across a wide surface of related questions. A cluster produces exactly that. A list of individually-chosen high-volume keywords produces isolated peaks with gaps between them, and the gaps are where fan-out goes looking.

There is a second effect that matters just as much. Sustained coverage of one subject associates your brand with that subject in a model’s representation. This is the entity payoff: you are not just answering questions, you are becoming the thing the model thinks of when the category comes up.

Anatomy of a cluster built for citation

Layer	Job	Typical count
Pillar	Define the subject, link outward to everything	1
Sub-topic pages	Answer one real question completely each	8 to 20
Comparison pages	X vs Y, alternatives, “which should I choose”	2 to 4
Decision pages	Cost, timelines, who it suits, when not to	2 to 4
Service page	The commercial destination the cluster supports	1

Scroll the diagram sideways to see all of it.

The last row is the one most content plans forget. A cluster with no commercial destination generates reading, not pipeline.

How to build one

1. Pick a topic you can plausibly own

Choose a subject narrow enough that fifteen pages constitutes genuine depth. “Marketing” is not a cluster. “Generative engine optimization” is. If you cannot imagine being one of the three best sources on the internet for it, narrow further.

2. Map the question space, not the keyword list

Keyword tools will show you a fraction of this. Supplement them with:

• Real prompts. Ask ChatGPT, Gemini and Perplexity your buyer’s question and record the follow-ups they suggest. These are fan-out queries.
• Sales and support conversations. The questions people actually ask before buying, which rarely have search volume and almost always have purchase intent.
• Competitor gaps. What the established sources in your space have not covered properly.
• The objections. “Is it worth it,” “can I do it myself,” “what does it cost,” “who is it wrong for.”

3. Give every page one job

One page, one question, answered completely. Two half-answers on one page produce a muddled passage that loses to a competitor’s dedicated one. This is the same logic as content chunking for AI, applied at the page level.

4. Interlink with descriptive anchors

Every supporting page links up to the pillar, the pillar links down to every supporting page, and closely related siblings link across. Use anchor text that describes the destination’s subject, since the link text contributes to how the target is understood. Avoid “click here” and avoid linking everything to everything, which dilutes the signal.

5. Route the cluster to a commercial page

Each supporting page should have a natural path to the service it supports. Not an aggressive CTA on every paragraph, one clear, relevant destination.

6. Cover the decision stage, not just the education stage

Pricing, comparisons, and honest “who this is wrong for” pages are disproportionately valuable, because validation queries are both the closest to purchase and the hardest to get cited for. Citation rates for validation queries ran at 11.3% in the AirOps data, against 18.3% for product discovery. Fewer sources clear that bar, which is precisely why clearing it is worth the effort.

7. Maintain it

A cluster with three stale pages loses credibility across the whole set. Re-date, re-verify statistics, and add pages as new sub-questions emerge. Fan-out surfaces change faster than editorial calendars.

What to measure

Cluster performance is not a per-page metric. Track:

• Coverage. What share of your mapped question space has a page.
• Citation breadth. How many distinct pages in the cluster get cited, not just the pillar.
• Presence rate across a prompt set built from the cluster’s questions, per measuring brand visibility in ChatGPT.
• Share of model against competitors in the category, per share of model.
• Assisted conversions into the service page the cluster supports.

Expect the pillar to underperform your expectations and the specific, unglamorous sub-pages to overperform. That is fan-out working as intended.

Where this leaves you

The old model was to find high-volume keywords and write the best page for each. That still helps you rank, and ranking still gates retrieval. But the queries actually deciding AI citations are mostly invented, low-volume, and unpredictable, so coverage beats precision. Build the cluster, cover the real question space including the awkward commercial questions, and interlink it so the whole set reinforces one entity.

If you want a cluster mapped against what AI engines are actually asking in your category, that planning is where our AI SEO service starts.