Topic Clusters and Pillar Pages in the AI Search Era

A topic cluster used to be a way to concentrate authority: one broad pillar page, several narrower pages pointing to it. The old payoff was rank. The new payoff is different, because an AI answer is built from several smaller searches, and a tidy cluster gives you a candidate for more than one of them.

This is written for content leads who know the SEO version already. It sets out what is different, walks through a sample cluster and flags the cases where the structure earns nothing. The only assumption is that you control your site's architecture.

How one question fans out

Google describes a technique it calls query fan-out, in which AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources. Walk through the effect. Someone types a broad question about handling vendor security reviews. The system turns it into separate lookups for definitions, evidence, roles, risks and cost. Each lookup pulls passages from matching pages, and one answer is composed from the strongest, with citations when the engine chooses to show them.

Consider two sites. The first holds one long article that brushes all five sub-topics, so every passage has to beat a specialist page. The second holds a pillar and five focused pages, so each sub-topic has a page written for exactly that lookup. The engine-side view is in query fan-out explained.

Which searches trigger summaries

Pew Research Center studied March 2025 browsing by 900 US adults. About 18% of searches produced an AI summary, but the rate climbed from 8% for one- or two-word searches to 53% for searches of ten or more words, and reached 60% for searches phrased as questions. That is one market and one month, so treat it as a snapshot.

The design consequence is simple. Name spokes after complete questions in buyer wording. "Which vendor risks justify a full review?" earns more than "Vendor risk" ever could.

Ranking is a weaker gate than it used to be

Ahrefs checked whether URLs cited in Google AI Overviews also ranked in ordinary results. Its March 2026 update, using about 863,000 keyword results and four million cited URLs, found roughly 38% of cited URLs in the top ten, against about 76% in July 2025, with a caveat that its parsing changed between the studies. If retrieval reaches beyond the top results, a focused spoke that answers a sub-question cleanly can be selected even when it does not rank for the head term. That is the strongest argument for clusters in this era, and it is an argument about passages, not about authority flowing to a pillar.

A sample cluster, content invented

Let the pillar be "vendor security reviews". It defines the process, lays out the stages and points to each spoke. Beneath it sit five pages, each tied to one question. Which documents should a vendor supply? A checklist whose first line names the minimum set. Who signs off, and when may they refuse? A role table opening with the approver and the escalation route. How long should a review take? A short explainer that gives a realistic range and what stretches it. What counts as proof? A comparison of certificates, audit reports and questionnaires that opens with the verdict. How is the decision recorded? A template listing the fields to keep.

Every spoke links up to the pillar, plus one sideways link to a close relative, and the pillar links down to all five. That is the complete structure. The GEO guide is built the same way.

Where clusters stop helping

A cluster supplies more eligible passages and nothing else. It does not repair a weak brand, thin evidence or blocked crawlers. Ahrefs found branded web mentions correlated with AI Overview visibility at 0.664 against 0.218 for backlinks across 75,000 brands, which is correlation, not proof of cause, but it suggests that how others describe you matters alongside how you structure pages.

Why we would build fewer spokes

Fan-out tempts teams into fifty pages, one per imaginable sub-question. We would not. Thin spokes dilute the cluster and create duplicate answers that compete with one another. Google's guide for generative search also says there is no requirement to break content into tiny pieces and recommends organising for readers, while another Google page warns that pages added with no value for users may breach its spam policy.

Our working limit is five to ten spokes per pillar. Each has a named owner, opens with an answer that stands alone and says something the pillar does not. Merge before adding. The price is that some buyer questions stay unanswered on your site, which we accept in exchange for pages you can keep accurate. We would reverse the limit if tracking by sub-question showed narrow, thin pages cited more than your fuller ones, so test with a prompt list before expanding, using the prompt research method.

What the pillar itself should do

The pillar is not a dumping ground for everything the spokes cover. Its job is to define the topic, show how the parts relate and send readers to the right spoke. A pillar that tries to answer every sub-question fully becomes the long, light-touch article the cluster was meant to replace. Keep it short enough to scan, and let the spokes carry the depth.

Old clusters versus new ones

The classic version aimed to hand authority to the pillar. Its spokes were keyword phrases that began with an introduction, and success meant the pillar's rank. The AI-era version aims to supply a strong passage per sub-question. Spokes carry full-question titles, begin with a direct answer, and are judged by citations for each sub-question as well as rank.

Quick audit for an existing cluster

1. Name the five topics buyers link to your company. If that is hard, clusters do not exist yet.
2. For each topic, count pages linking both ways with the pillar. One or none signals a gap.
3. Read the opening sixty words of each spoke and rewrite any that merely introduce the subject.
4. Look for pairs answering one question. Merge them, and redirect the weaker address.
5. Scan for stale dates. A spoke quoting last year's figures needs refreshing or retiring.

Upkeep

Treat the cluster as living. Each quarter, update figures, fix dead links and add a spoke once sales has heard a new question three times. Remove those that no longer fit. Old pages cost more than before, since a system retrieving a passage may repeat a stale fact as if it were current. Finding decaying content shows which pages to check first.

To start, pick one pillar, draft ten genuine questions and publish the five that matter most. Ask a stranger whether each opening makes sense alone, then watch which spokes are cited. When one is skipped, the structure audit usually explains it.