How I structure topic clusters for SEO and AI visibility

A cluster that works in traditional search can fail in generated answers. The difference lies in the retrieval unit.

Disclosure: I am a Semrush Ambassador, and this article describes Semrush features. The working criteria described here hold with any equivalent tool, and the passages describing the product are marked with source and date checked.

Most of the topic clusters I audit were built from a keyword spreadsheet. Someone exported search volume, grouped by string similarity and turned each group into a brief. The result works reasonably well while the search system returns a list of links, because the whole page competes as a unit. Systems that generate answers work at another scale. They retrieve passages, assess whether that passage is enough to answer, and compose the answer from different sources. The unit that competes becomes the block. A cluster designed for the first logic tends to have good pages with fragile blocks, a bit like a beautiful house where every room depends on the hallway to make sense.

What I change in the cluster method

I start from the entity. The keyword comes later, as a symptom of an intent. Before defining any brief, I write the definition of the cluster’s central entity in one sentence that can be copied and remain true out of context. That sentence becomes the pattern every page in the cluster repeats, varying in form and never in meaning.

Then I map the intents orbiting that entity, separated into four tiers: definition, comparison, execution and decision. Each tier becomes a page, and each page needs to answer that tier’s question in a self-sufficient way in the first paragraphs. To read each intent, I use the real results for the query next to the tool’s label, as I describe in search intent without automatic classification.

With one page per tier, the minimum cluster has four pages per central entity, and it grows when a tier calls for more than one answer, such as comparisons with different competitors. Internal linking follows the same logic: every page points to the definition page with the same anchor, and the definition page points to the others by the question each one answers. That is how the relationship between the pieces becomes legible to readers and to the system doing the retrieval.

The third step is the one that sparks the most debate in teams: I write each section as if it were going to be retrieved alone. That means repeating the subject instead of using a pronoun, naming the entity instead of writing “the tool”, and closing the section with a statement that survives being cut out.

The cluster step by step in Semrush

The sequence below is the one I run, in this order, and the order matters because each step narrows the next. Starting with the wrong tool produces a cluster that covers territory without answering any decision.

• Keyword Gap to draw the territory. I compare the client’s domain with three competitors fighting for the same results, even if they are not on the commercial competitor list, and look at where coverage is missing by intent.
• Prompt Research to bring in the language of people asking a generative system, which tends to be longer and more contextual than a query typed into traditional search.
• Site Audit to make sure the cluster’s pages can be crawled and rendered, because a cluster that is flawless in content and blocked at template level does not exist for any system.
• Position Tracking to follow the definition tier, which holds up the other three when it works.

The point of attention in the sequence is not closing the list before checking how the generated answer treats that territory. That check is what reveals when the subject is already occupied by a source that did not show up in the competitive analysis. What leads a source to take that place is in the three conditions to become an AI reference source.

How I know the topic cluster worked

The first signal shows up in the definition tier, when the anchor page starts being cited in answers that do not use the client’s exact name. The second signal is description consistency, meaning the system starts describing the brand with the same attributes the anchor page established.

The third signal is the slowest and the most valuable. The execution pages start getting qualified traffic from people who already understood the definition somewhere else. At that point the cluster starts working as a system, with pieces that hold each other up.