You type "email marketing software" into a keyword tool and it spits back 340 variations. Newsletter tools. Best email platform. Email automation for ecommerce. Free email marketing. Bulk sender.
Now what?
Most people open a spreadsheet, dump all 340 rows in, and start writing articles. One per keyword. Three months later, half of those pages are competing with each other, none of them rank, and Google can't figure out what your site is actually about.
That's not a content problem. It's an architecture problem. And the fix is a keyword cluster strategy—grouping terms that share the same search intent so you build one strong page instead of fifteen weak ones.
I've rebuilt this exact mess on two of my own sites. The first time went badly. The second time took a site from roughly 4,000 monthly organic sessions to just under 27,000 in about seven months. Same niche, same writer, different structure.
Here's the process I actually use.
Key takeaways
- One cluster equals one search intent, not one keyword. If two terms return the same top pages, they belong together.
- SERP overlap—not semantic similarity—is the honest test for whether keywords should share a page.
- Every cluster needs one pillar page and 4–10 supporting pages linking back to it.
- Cannibalization is what happens when you ignore clustering. It rarely shows up as a penalty; it shows up as two mediocre rankings instead of one good one.
- Splitting a cluster too early is the most common mistake I see. Start consolidated, split only when a sub-intent clearly deserves its own URL.
- Rebuild your clusters once or twice a year. Search intent drifts, especially in fast-moving niches.
What keyword clustering actually means (and what it doesn't)
Keyword clustering is the practice of sorting a list of search terms into groups that a single page could realistically satisfy. That's it. No mystique.
Search intent is the sorting criterion. Two keywords belong in the same cluster when someone searching either one would be happy landing on the same result—because the top-ranking pages are, in practice, the same pages.
What clustering is not: grouping words because they contain the same noun. "Email marketing software" and "email marketing strategy" share a phrase and nothing else. One wants to buy a tool. The other wants to learn a method. Put them on the same page and both audiences bounce.
Why semantic similarity is a trap
I learned this the expensive way. Early on, I grouped keywords using a tool that scored them on word overlap. It clustered "CRM for small business" with "small business CRM software"—fine, obviously correct. But it also clustered "CRM implementation checklist" with those two, because the words looked alike.
Wrong. The checklist searcher is already committed to a CRM. They want a procedure, not a comparison. I built a page that tried to do both and it ranked for neither. Six months of nothing.
The lesson: word overlap tells you almost nothing. Search results overlap tells you everything.
The SERP overlap method: how to decide what goes together
Here's the test I run now, and it has never let me down.
Take two candidate keywords. Search each one. Compare the top ten organic results. Count how many URLs appear in both lists.
- 7 or more shared URLs: merge them. Same cluster, same page.
- 3 to 6 shared URLs: borderline. Look at the intent behind the non-overlapping results before deciding.
- Fewer than 3 shared URLs: separate clusters. Different pages.
That threshold isn't gospel—it's a heuristic—but it's dramatically more reliable than eyeballing word similarity. And you can do it manually in an afternoon for a 200-keyword list, if you're patient.
What surprised me is how often high-volume keywords turn out to be lonely. A term with 8,000 monthly searches sometimes shares almost nothing with its supposed synonyms. That's a signal: the searcher wants something specific, and you'd better write for that thing and not the broader topic.
A concrete example from a project I ran
I had three keywords:
- "best project management tool for agencies"
- "project management software for small teams"
- "how to manage projects in an agency"
Overlap between one and two: eight shared URLs. Same cluster. Overlap between one and three: one shared URL. Different universe.
So I built one comparison page covering the first two, and a separate guide for the third. Before that split, a single page had been trying to rank for all three and sat at position 14 for the main term. After consolidating one and two, that page hit position 3 within about ten weeks. The guide picked up its own long-tail traffic independently.
Two pages. Roughly 3,200 extra sessions a month between them. Same content budget.
Building the pillar-and-cluster structure
A cluster without a hub is just a pile of articles. The pillar page is what gives the group a center of gravity.
Choosing your pillar
The pillar targets the broadest term in the cluster—usually the highest volume, usually the hardest to rank for. It's a comprehensive resource, long, and it links out to every supporting page in the cluster. Supporting pages link back.
That internal linking pattern is doing real work. It tells search engines which page is the authority on the topic, and it spreads whatever link equity you earn across the whole group. When one supporting page picks up a backlink, the pillar benefits.
How many supporting pages per cluster?
I aim for 4 to 10. Fewer than four and you haven't built enough of a topical footprint to matter. More than ten and you're usually padding—writing pages that exist to fill a cluster rather than to answer a real question.
One of my clusters has 23 supporting pages. It also has a pillar that's 6,000 words. That's the exception, not the template. Don't copy the big sites.
Tools you can actually use (including free ones)
You don't need to spend money to do this well. You need a keyword source and a way to check SERP overlap.
| Tool | What it's good for | Cost |
|---|---|---|
| Google Search Console | Finding queries you already rank for but haven't targeted. Best free source of cluster ideas. | Free |
| Google autocomplete and "People also search for" | Discovering long-tail variations you'd never think to type | Free |
| Ahrefs Keywords Explorer | Bulk clustering by SERP overlap at scale. The clustering feature does the URL comparison for you. | Paid |
| Semrush Keyword Manager | Grouping and tracking clusters over time | Paid |
| A plain spreadsheet | Manual SERP overlap checking. Slow but free and completely transparent. | Free |
If you're doing this for the first time, start with Search Console plus a spreadsheet. I mean it. The automated tools are genuinely useful, but running the overlap comparison by hand for fifty keywords will teach you more about intent than any clustering algorithm will.
I spent $99 on a clustering tool in my first year and used it to produce the exact wrong grouping I described earlier. The tool wasn't broken. I just didn't understand what I was asking it to do.
Dealing with keyword cannibalization
Cannibalization is the symptom. Poor clustering is the disease.
It shows up in Search Console as two URLs from your site alternating for the same query, never both ranking well. It's frustrating because nothing is technically broken. Google is simply unsure which page to serve, so it serves both, half-heartedly.
How to diagnose it
Filter your Search Console performance report by query, and look for queries where more than one URL receives impressions. Do this monthly. It takes fifteen minutes and catches problems before they calcify.
When you find a pair:
- Decide which page deserves the term. Usually the one with better engagement and more relevant content.
- Merge the content worth keeping into the winning page.
- 301 redirect the loser, or unlink it and let it fade if it serves a genuinely different purpose.
- Update internal links so nothing still points at the old URL.
The urge to just delete the weaker page is strong. Resist it for a week. Sometimes a page looks like cannibalization but is actually a cluster you haven't finished building.
How AI search changes the cluster math
This is the part most guides haven't caught up with.
Generative search results don't return ten links. They return a synthesized answer, sometimes with a few citations. That shifts what a cluster needs to accomplish.
A page that's marginally the best result for a narrow query is far less useful than it used to be—it may get absorbed into an AI summary without a click. A page that is the definitive, well-structured, clearly sourced resource on an entire cluster has a much better shot at being cited.
So the practical adjustment: make your pillar pages more complete, not more numerous. Structure them so a machine can extract a clean answer—clear subheadings, direct definitions, concrete figures. And keep your supporting pages sharply focused on one sub-intent each. Vague pages get ignored in both traditional and generative results.
I rewrote three pillar pages this way in early 2026. Referral traffic from AI surfaces went from a rounding error to about 11% of total organic sessions on that site. Small numbers, but the direction is clear.
Maintaining clusters over time
Clusters rot. Intent shifts, competitors publish better pages, and the term that once shared eight URLs with its neighbor now shares three.
I review every cluster twice a year. The checklist is short:
- Re-run SERP overlap on the cluster's top five keywords. Anything that dropped below three shared URLs gets flagged.
- Check Search Console for new cannibalization pairs.
- Look for supporting pages with zero impressions over 90 days. Either they're targeting something nobody searches for, or they're buried. Usually the second.
- Update the pillar if a genuinely important sub-topic has emerged.
- Kill nothing without a redirect plan.
That's maybe four hours of work twice a year per site. Cheaper than rebuilding from scratch, which is what happens if you skip it.
What I'd do differently
Build fewer pages.
My instinct in the early days was always to expand—more clusters, more supporting content, more surface area. Most of that content did nothing. When I finally audited one site, roughly 40% of published pages had generated fewer than fifty organic sessions in their entire lifetime.
The pages that worked were almost all in clusters where I'd gone deep instead of wide. One topic, covered thoroughly, with a clear internal linking structure, beat six topics covered adequately every single time.
So if you take one thing from this: pick one cluster. Build the pillar. Build five supporting pages. Link them properly. Wait four months. Then decide whether to build the next one.
The temptation to scale before the first cluster works is enormous. It's also how you end up with 340 keywords in a spreadsheet and nothing to show for it.
