Building Intent Clusters Around Real Information Needs

Ten differently worded prompts can be one question wearing ten different outfits. "What CRM should a small team use," "best CRM for under 10 people," "CRM recommendations for a startup," and "which CRM is easiest for a small business" are not four separate content opportunities. They are one intent, phrased four ways by four different people who never coordinated with each other. Building a separate page for each wording variation produces four thin, redundant pages competing with each other instead of one strong page serving all four prompts well. This sub-chapter is about the skill of seeing past the wording to the need underneath it, and building content around that need instead of around any single phrasing of it.

What an Intent Cluster Is

An intent cluster is a group of prompts, regardless of how differently they are worded, that share the same underlying information need. The wording is surface variation. The need is the thing that actually matters for content planning.

This distinction directly answers the instinct that produces bloated, redundant content: the instinct to see ten different prompt wordings and conclude there must be ten different content opportunities. There usually are not. Real people ask the same underlying question in dozens of phrasings, shaped by their own vocabulary, their platform of choice, and how they happen to be thinking about the problem at that moment. An intent cluster is the practitioner’s tool for seeing past that surface noise to the actual, smaller set of real needs sitting underneath a large, messy set of prompt variations.

The Underlying-Need Test for Grouping Prompts

The practical test for whether two prompts belong in the same cluster is direct: could one well-built piece of content resolve both of them completely, without either one feeling like an afterthought or a mismatch.

If the answer is yes, they belong together regardless of how different the wording looks on the surface. “What CRM should a small team use” and “best CRM for under 10 people” pass this test easily. Both are asking, in different words, for a recommendation appropriate to small team size. A single, well-built comparison and recommendation page resolves both prompts fully. If the answer is no, if serving one prompt well would leave the other one genuinely unaddressed, they belong in separate clusters even if the surface wording looks similar. “What CRM should a small team use” and “how do I migrate CRM data from Excel” share a topic area but not an underlying need, and forcing them into one cluster produces content that serves neither prompt particularly well.

The Axes Clusters Can Be Built Along

Intent clusters can form along several distinct axes, and a mature prompt research practice tracks more than one at once rather than defaulting to topic as the only organizing principle.

Topic is the most obvious axis: prompts about the same subject matter. Funnel stage is a second axis: prompts from someone just learning a topic exists cluster differently than prompts from someone actively comparing final options. Task is a third: prompts oriented around completing a specific action cluster separately from prompts oriented around understanding a concept, even within the same topic. Risk is a fourth: prompts carrying visible caution or high-stakes framing, common in regulated or expensive categories, often need a different treatment than low-stakes versions of a similar question. Geography is a fifth, relevant wherever local variation genuinely changes the answer. Sophistication level is a sixth: a beginner’s phrasing and an expert’s phrasing of a related question frequently need different depth and vocabulary even when the topic overlaps heavily. A single topic area often needs to be sliced along two or three of these axes simultaneously to produce clusters that are genuinely coherent.

One Cluster, One Asset

The practical payoff of correct clustering is a direct one-to-one relationship, in most cases, between a cluster and a single strong asset built to resolve it fully, rather than a scattering of thin pages each addressing one prompt wording.

This is not a rule against ever having more than one page on a related subject. It is a rule against building redundant pages that exist only because someone found a different way to phrase the same question. A single well-built asset, structured with the extractable-block discipline covered in Chapter 4.5, can serve every prompt wording within a genuine cluster, because the underlying need is what the content is built to resolve, not any one surface phrasing of it. This is also more efficient from a production standpoint: one strong asset built well outperforms four thin assets built quickly, on essentially every dimension that matters, including the generative retrieval outcomes this framework cares about most.

When a Cluster Is Actually Two Clusters

The most common clustering mistake runs in the opposite direction from over-fragmentation: grouping two genuinely different needs into one cluster because they share surface topic language, then building one asset that serves neither need particularly well.

A cluster that tries to serve both “what is GSO” (definitional, early-stage curiosity) and “how much does GSO cost to implement” (evaluative, late-stage decision-making) under one roof produces a page pulled in two directions, too basic for the second prompt and too commercially focused for the first. The underlying-need test catches this directly: if one well-built asset genuinely cannot resolve both prompts without shortchanging one of them, the cluster needs to split. This usually surfaces during content planning rather than during initial research, which is why revisiting cluster boundaries as content gets built, rather than treating the first-pass clustering as final, is a normal and expected part of the process.

Clustering as an Ongoing Practice

Clustering is not a one-time exercise performed at the start of a content program and then left untouched. New prompt language surfaces continuously as research continues, as Chapter 7.6 covers, and existing clusters need to absorb, or sometimes split to accommodate, that new language over time.

A cluster that looked coherent with twenty observed prompts can reveal a genuine internal split once fifty more prompts have been gathered and a pattern within the cluster becomes visible that wasn’t apparent with the smaller sample. Treating cluster boundaries as provisional, subject to revision as more prompt data comes in, produces a more accurate map over time than treating the first clustering pass as a permanent structure. The output of this ongoing process feeds directly into the coverage assessment in Chapter 7.4, which depends on clusters being genuinely coherent to measure coverage against.

Clustering by Need, Not by Convenience

Michael Rubinstein treats intent clustering as the single highest-leverage skill in prompt research, because getting it wrong in either direction, fragmenting one need into many thin pages or merging several needs into one unfocused page, produces content that looks reasonable in a production tracker and underperforms consistently against real prompts.

ScribePress applies the underlying-need test systematically before any content brief gets written, specifically to catch both failure directions before they turn into published pages that need to be reworked later.

Learn more about the work behind this framework at michael-rubinstein.com.

Put the framework to work