Prompt clusters are replacing keyword clusters.

For about a decade, the keyword cluster was the atom of content planning. You picked a head term, gathered its long-tail variants, mapped them to a pillar page and a handful of supporting articles, and shipped. It worked because the search box rewarded it. Someone typed "project management software," Google matched documents that used those words often and well, and the page that covered the cluster most thoroughly tended to win.

That machine still runs, but a growing share of searches no longer touch it. When a query lands in Google's AI Mode or triggers an AI Overview, the system does not hand back ten blue links for you to compete over. It reads the question, breaks it into sub-questions, retrieves passages from across many sources, and assembles an answer. The unit it is answering is not a keyword anymore. It is a full request, usually a long and specific one, and the content plans that still start from a keyword list are optimizing for a match that increasingly does not happen.

The practitioners paying attention have started planning in a different unit. Instead of a keyword cluster, they build a prompt cluster, a group of conversational, context-rich queries that trace how a real person talks to an AI assistant across a decision. This guide walks through what that shift actually means and how to build a prompt cluster from scratch, with worked examples you can copy.

What a keyword cluster was good at

A keyword cluster is a set of related search terms grouped by topic and mapped to content. You take a seed like "keyword clustering," pull the variants people search around it, and organize them into a pillar plus supporting posts. Semrush still shows the demand clearly, "keyword clustering" runs about 1,600 US searches a month and "topic clusters" around 1,000, so this is not a dead idea. It is the backbone of how most sites are organized, and for classic blue-link search it holds up well.

The logic was lexical at its core. Group the terms, cover them densely, interlink the pages, and you signal topical authority to a ranking system that mostly matched words to words. A page targeting "best running shoes for flat feet" earned its place by using that phrase and its neighbors more thoroughly than the competition. You could see the whole game on a spreadsheet, one row per keyword, volume and difficulty in the next columns, a target URL at the end.

The weakness only showed up once the answer layer arrived. A keyword cluster tells you which words people type. It says very little about the shape of the question behind those words, the context the person is carrying, or the three follow-ups they will ask next. When Google answered with links, that gap did not matter, because the person did the follow-up work themselves by clicking around. When Google answers with a synthesized paragraph, the gap becomes the whole problem.

What AI Mode changed about the query itself

The first thing that changed is length. Queries typed into AI Mode and other conversational search surfaces run far longer than the two or three words people fed the old box, with several analyses putting the average at roughly three times the length of a traditional search. Nobody types "running shoes" into a chat assistant. They type "I overpronate and my knees hurt after 5k, what shoes should I look at under a hundred dollars." The query carries a condition, a symptom, a distance, and a budget before it ever names a product category.

The second change is what the engine does with that longer query. Rather than retrieve pages for the string as a whole, it fans the query out into constituent questions, retrieves passages that answer each one, and stitches them into a single response. Google has described this fan-out behavior directly, and it means a single visible search can pull from a dozen underlying sub-queries you never see. Your content is no longer competing to match one phrase. It is competing to be the passage the model selects for one of those hidden sub-questions.

The third change is the one that should reset how you plan. The sources feeding AI answers are drifting away from the classic top ten. One widely cited data set found that the share of Google AI Overview citations coming from pages that also rank in the organic top ten fell from roughly 76% to 38% over about a year and a half. Ranking on page one and getting cited in the answer used to be nearly the same achievement. They are now separate outcomes with separate requirements, and a keyword cluster only optimizes for the first one.

What a prompt cluster actually is

A prompt cluster starts where a keyword cluster stops. Instead of a head term and its lexical variants, you build around a core topic plus the layers that surround how someone actually asks about it, the context they bring, the modifiers that narrow their intent, the follow-up questions that come next, and the intent they imply without stating. The output is not a list of phrases to hit. It is a map of the conversation, structured the way an AI assistant assembles an answer rather than the way a search box matches a string.

Picture the difference on the same topic. A keyword cluster for a payroll tool might be "payroll software," "payroll software for small business," "best payroll software," "payroll software pricing," and a few more, all sitting flat on a spreadsheet. The prompt cluster for the same tool looks like a decision unfolding. It opens with "what's the easiest way to run payroll for a 12-person company," moves to "do I need payroll software if I only have contractors," picks up modifiers like "cheap," "with direct deposit," and "that files taxes for me," then runs into follow-ups such as "how long does it take to switch payroll providers" and "will my accountant still have access." The implied intent underneath all of it is a small business owner who is nervous about getting payroll wrong and wants reassurance as much as features.

Content built for that map reads differently and gets retrieved differently. Each real question becomes a passage the model can lift cleanly, phrased the way the user asked it, answered in the first line before you elaborate. You are still covering a topic in full, so the old instinct is not wasted. You are just organizing coverage around the questions an assistant will decompose the request into, which is why a prompt cluster tends to win citations that a lexically identical keyword cluster misses.

How to build a prompt cluster from scratch

Start with the core topic and one real user, not a keyword. Write down who is asking and what they are actually trying to accomplish. For a project management tool, that might be an operations lead at a 40-person agency who is tired of tracking work in spreadsheets and needs to convince a skeptical team to switch. Everything downstream gets sharper when you anchor it to a person with a situation rather than a term with a volume.

Next, gather the real prompts. This is field work, not brainstorming. Pull the actual long queries from your Search Console pages that already earn impressions in AI surfaces, read the "People also ask" and the follow-up chips inside AI Mode for your topic, and open a fresh chat with ChatGPT, Gemini, and Claude to ask your core question the way a customer would, then watch what those assistants suggest asking next. You are collecting the sentences people say, in their words, with their context attached.

Then sort those prompts into the layers of the cluster. Put the core question at the center. Around it, group the context variants, someone asking as a startup founder versus an enterprise buyer wants different framing of the same answer. Layer in the modifiers that narrow intent, the "for remote teams," "free," "without a credit card," "for non-technical users" qualifiers. Add the follow-up questions that come after the first answer, since those are often where competitors have written nothing. Finally, name the implied intent behind the set, the unstated worry or goal, because that is what your framing and your examples should speak to even when no query says it out loud.

Now map each prompt to a passage, not a page. A prompt cluster usually still lives on a pillar page and a few supporting articles, so the structure is familiar, but the internal organization changes. Each question becomes a section or a sub-heading phrased as the question, with a direct answer in the opening sentence and the supporting detail after it. That direct-answer-first shape is what makes a passage liftable, and it is the difference between being the page that ranks and being the passage that gets quoted.

Close the loop by writing to be retrieved, then checking whether you were. Give every important claim a concrete number, a named source, or a specific example, because those are the elements AI answers pull most reliably. After the content is live, run your core prompts back through the AI assistants and note which of your passages get cited and which questions still return a competitor. The gaps become your next round of edits. A prompt cluster is not a one-time deliverable, it is a living map you refine against the answers the models are actually giving.

Three WORKING examples

Take a B2B SaaS example first. A time-tracking tool would once have targeted the keyword cluster "time tracking software," "time tracking app," "employee time tracking," "time tracking for freelancers." The prompt cluster reframes the same territory as a conversation. The core prompt becomes "how do I track billable hours across a small team without annoying everyone," the modifiers pull in "that integrates with QuickBooks," "with a mobile app," and "that doesn't feel like surveillance," and the follow-ups surface "how do I get my team to actually log their time" and "can I bill clients directly from tracked hours." That last cluster of worries is where most competitor content is silent, so a page that answers them in liftable passages can win citations even from a lower ranking position.

Now an ecommerce example. The old cluster for a running-shoe retailer was "running shoes," "trail running shoes," "running shoes for flat feet," "best running shoes 2026." The prompt cluster begins from the way a runner actually describes the problem, "what running shoes help if I overpronate and get knee pain on longer runs," then layers modifiers like "under a hundred dollars," "for wide feet," and "good for both road and treadmill," and follows with "how often should I replace running shoes" and "should I size up for long distance." A comparison table of models against pronation type and price earns its place here, because that is genuinely tabular and a paragraph would handle it worse. Everything else stays prose that answers a specific question in its first line.

For a local service example, take an Indonesian fertility clinic optimizing for both Google and AI answers. The keyword cluster would have been "klinik fertilitas," "program bayi tabung," "biaya IVF Jakarta." The prompt cluster maps the emotional decision a couple is actually working through, opening with "how do we know when it's time to see a fertility specialist," carrying context like "we've been trying for a year" and "we're in our late thirties," picking up modifiers around cost, success rates, and clinic location, then running into the follow-ups that keep people awake, "how painful is the IVF process," "what are the odds it works on the first cycle," and "what happens if the first round fails." Answering those directly, with real figures and clear sourcing, is both better content and better AI-retrieval material, and it speaks to the implied intent, a couple looking for honesty and reassurance, that no single keyword ever captured.

The content formats that actually get cited

Format is not cosmetic in AI search, it changes your odds of being selected. Analysis from Wix's AI Search Lab, which looked at more than a million citations, found that listicle-style pages pulled roughly a 21.9% share of AI citations against about 16.7% for standard articles, because a clean list gives the model discrete, self-contained items it can lift without untangling your prose. That does not mean turn every page into a listicle, it means that where the material is genuinely a set of options or steps, structuring it as a real list makes each item retrievable on its own.

Data does even more work. The foundational academic study on this, the Princeton-led Generative Engine Optimization paper, tested content changes across generative engines and found that adding statistics, citing sources, and including direct quotations produced the largest visibility gains, lifting a page's presence in AI answers by up to around 40%. The mechanism is intuitive once you see it. A model assembling an answer prefers a passage that states "cycle success rates average 31% per attempt for patients under 35" over one that says success rates "vary widely," because the specific claim is quotable and the vague one is not.

The third format lever is the direct-answer structure inside the page. Lead each section with the answer to the question that heads it, then support it, rather than building up to the point across three paragraphs. AI systems retrieve at the passage level, so a section that front-loads its answer gives the model a clean unit to cite, while a section that buries its conclusion at the bottom asks the model to do work it will usually skip. This is the single cheapest change most pages can make, and it costs nothing but the discipline of reordering what you already wrote.

Your high-impression, low-CTR pages are the fastest win

Before you build anything new, look at the pages you already have that get shown constantly and clicked rarely. In Search Console, sort your pages by impressions and find the ones with high impressions and a click-through rate near the floor. In the AI-search era that pattern often means the page is being surfaced as a source, read by the answer engine, and used to inform a response the user never clicks through to read. The impressions are real. The value is leaking because the page was written to rank, not to be cited.

These pages are the fastest win because the hard part is already done. Google already considers them relevant enough to show, so you are not fighting for visibility from zero, you are reformatting content the system already trusts. Take one of these pages and rebuild it around its prompt cluster. Rewrite the headings as the actual questions people ask, move the direct answer to the top of each section, add the specific numbers and named sources you were missing, and break genuine option-sets into clean lists. You are not writing more, you are restructuring what is there into liftable units.

The reason this works is that reformatting matches the page to how it is already being used. A page that earns thousands of impressions in AI surfaces is being retrieved and read by the model, and small structural changes can move it from "read and paraphrased anonymously" to "quoted and attributed," which is where the brand exposure and the referral traffic actually live. Run this pass across your top twenty high-impression, low-CTR URLs before you commission a single new article, because the return per hour of work is higher than almost anything else on the roadmap.

How to know it is working

The old scoreboard does not fully capture this, so add to it rather than throwing it out. Keep tracking rankings and organic clicks, since classic search still sends real traffic and a page that ranks well is more likely to be retrieved. But layer in citation tracking, the practice of running your core prompts through ChatGPT, Gemini, Claude, and Perplexity on a schedule and recording where your brand and your pages show up in the answers. That is the metric a prompt cluster is actually built to move, and it is invisible on a rankings report.

Watch the shape of your Search Console data too. A rising impression count paired with a flat or falling click-through rate is not necessarily a failure anymore, it can be the signature of a page being used as an AI source. Read it alongside your citation tracking rather than in isolation. When impressions climb, citations appear, and the questions you built passages for start returning your name in the answer, the cluster is doing its job even if a raw traffic chart looks unremarkable.

Where this leaves keyword research

None of this retires keyword research. Search volume still tells you which topics carry real demand, difficulty still tells you where you can realistically compete, and the keyword data is exactly what told us "prompt clusters" as a phrase has almost no search volume yet while the intent around it, AI Mode, query fan-out, generative and answer engine optimization, is climbing fast. The keyword layer is how you pick which conversations are worth mapping. The prompt cluster is how you win them once you have chosen.

The honest way to think about it is a promotion rather than a replacement. Keywords move from being the thing you optimize for to being the demand signal that points you at the right prompt clusters to build. Sites that make that move are structuring content around how AI assembles answers, and they are the ones getting cited while everyone else is still filling in a spreadsheet of phrases and wondering why their page-one rankings stopped translating into traffic.