Technical SEO

Keyword Grouping Software: How to Choose and Test One

· · 14 min read

Search for keyword grouping software and you get a list of tools, ranked. Four of the pages currently ranking for that term are published by a vendor that places itself first. Not one of them shows you how to check the ranking.

That is the gap this article fills. What follows is an acceptance test you can run in about an hour on your own keyword list: build an answer key, score what each tool returns against it, work out what a usable cluster actually costs, and check whether the same list produces the same groups a week later. At the end there is a scorecard.

You will not find a ranked list of tools here. You will find a way to produce your own.

Key takeaways

  • Keyword grouping software sorts a keyword list into groups that a single page can serve — and Ahrefs’ study of 3 million queries found the average top-ranking page already does this for nearly 1,000 keywords.
  • A vendor’s accuracy claim is unverifiable until you have an answer key of your own. Build one before you start a trial.
  • Score two things separately: whether anything foreign got into a group, and whether anything that belonged got left out.
  • The overlap threshold is a dial you set, not a property of the tool. In my own test, seven near-identical search terms formed seven groups at one setting and two at another.
  • Compare tools on cost per usable cluster, never on the headline price per thousand keywords.

What keyword grouping software actually does

Keyword grouping software takes a flat list of search terms and sorts it into groups, where each group is meant to represent one page. The premise is sound: pages already rank for far more than the term you wrote them for. Ahrefs’ study of 3 million queries, cited above, found that the average number one result also ranks in the top ten for nearly a thousand other relevant keywords. You are not choosing one keyword per page. You are choosing which hundred belong together.

Get that wrong in the other direction and you pay for it. Publish two pages aimed at the same intent and you have two sets of weaker signals instead of one strong set — which is the reason Google Search Central describes canonicalisation as a way to “consolidate the signals they have for the individual URLs (such as links to them) into a single, preferred URL”. Grouping is how you avoid needing that rescue in the first place.

Tools do the sorting in one of two ways, and the distinction matters more than any feature list.

Word similarity. The tool compares the keywords themselves — shared words, stems, or a semantic embedding of the phrase. It is fast, it costs almost nothing to run, and it has no idea what Google thinks. “Cheap flights to Rome” and “Rome flight deals” look similar to this method whether or not they return the same results.

Search-result overlap. The tool fetches live results for every keyword and groups two terms when enough of the same URLs appear for both. This asks Google directly rather than inferring from the words, which is why it costs more: somebody has to pay for all those result pages.

Most of the frustration people report with grouping tools comes from running a word-similarity tool while expecting search-result behaviour. Before you evaluate anything, find out which method you are buying.

Build a gold-standard set before you trial anything

Every keyword grouping software vendor claims accuracy. None of those claims mean anything to you until you have something to measure them against.

So build an answer key. Take 100 to 150 keywords from a topic you know thoroughly — your own product area, ideally, where you already know which terms are the same question in different clothes. Group them by hand. Do it once, save it, and never change it to flatter a tool’s output. This is exactly how clustering is evaluated in practice: scikit-learn’s clustering evaluation documentation frames every supervised metric as a comparison between “the ground truth class assignments” and “our clustering algorithm assignments of the same samples”. Your hand-grouping is the ground truth.

Three rules make the set worth having.

Include the awkward cases deliberately. A set of obviously distinct topics will give every tool a perfect score and tell you nothing. You want the near-misses — the term that could sit in two groups, the modifier that changes intent, the plural that returns a different page type. Those are where tools differ.

Write down why, not only what. For each group, note the page you would build. If you cannot name the page, the group is wrong and you have found a flaw in your answer key before it corrupts the test.

Decide your rule for one-offs in advance. Some keywords belong to no group. Decide now whether a singleton is a legitimate answer or a failure, because tools differ wildly in how many they produce, and you do not want to be making that judgement while looking at the results.

Budget about forty minutes for this. It is the only part of the test you do once — the set then grades every tool you ever evaluate, including the one you eventually buy, a year from now, when you want to know whether it has quietly changed. If you want a no-cost first pass to react against, our free keyword clustering tool will group a list without a trial sign-up.

Score the output: what got in, what got left out

Run your answer key through a piece of keyword grouping software and you get back its grouping. Now you need a number.

The temptation is to count groups that match exactly. Resist it — exact matches are rare and a single misplaced keyword destroys an otherwise good group. Score pairs instead. Take every possible pair of keywords in your set and ask two questions: did my answer key put these two together, and did the tool?

That gives you two separate failures, and they are not the same problem:

  • Contamination. The tool put two keywords together that your key keeps apart. You get a group containing something foreign.
  • Splitting. Your key puts two keywords together and the tool separated them. A group that should exist has been cut in half.

These map onto the two objectives scikit-learn’s clustering documentation, linked above, uses to score a clustering: homogeneity, where “each cluster contains only members of a single class”, and completeness, where “all members of a given class are assigned to the same cluster”. You need both numbers, because either one alone is trivially gamed. A tool that returns every keyword in one enormous group is perfectly complete and useless. A tool that returns 150 groups of one is perfectly homogeneous and equally useless.

A spreadsheet handles the arithmetic — for 150 keywords there are 11,175 pairs, which is nothing for a formula and impossible by eye. Count the pairs you and the tool both grouped together, the pairs it merged that you did not, and the pairs it split that you did not. Two fractions fall out.

My working bar: I will run a tool unattended if it clears roughly 0.80 on both. Between 0.60 and 0.80 it is useful with a human reviewing every group before anything gets briefed. Below 0.60 on either number, the review takes longer than grouping by hand.

Compare that to what the market tells you. One ranking page states that modern AI models “achieve impressive accuracy in intent classification, often exceeding 90% correctness when distinguishing between informational, commercial, transactional, and navigational queries” — with no study, no methodology and no link behind it. That figure might be right. You have no way of knowing, and neither does the person who published it. An hour with your own answer key replaces every claim of that kind.

Turn the threshold dial before you judge any tool

Here is the finding that changes how you run the whole test: most of what you will attribute to tool quality is actually a setting.

Search-result tools merge two keywords when they share enough of the same ranking URLs. How many is “enough” is yours to choose. SE Ranking’s own documentation is explicit about it — you “define the minimum number of matching URLs and clustering method”, and “the higher the accuracy the less search queries will be added to one group”. Tighten the dial and you get many small, confident groups. Loosen it and you get few large, speculative ones. Same tool, same list, different answers.

I measured this while researching this article. I took seven terms that any writer would call synonyms — keyword grouping software, keyword grouping tool, keyword grouper, keyword grouping, keyword clustering software, keyword clustering tool, keyword clustering tools — and pulled live US results for each. Across all 21 possible pairs, not one shared the three ranking URLs that most search-result tools require before merging two keywords. Twelve of the 21 pairs shared no URLs at all. The strongest pair overlapped by two results out of ten.

Then I moved the dial. At a 0.15 overlap threshold, those seven terms stayed as seven separate groups. At 0.05, they collapsed into exactly two: everything with “grouping” in it, and everything with “clustering” in it. Nothing about the tool changed between those two runs. Only the number did.

Two things follow. First, test every tool at more than one setting, or you are grading the default rather than the software. Second, words are a poor guide to what Google treats as one topic — seven terms that read as synonyms returned seven essentially different result sets. If you want the mechanics of how overlap scoring works, I have written up the SERP overlap clustering method separately.

Price it per usable cluster, not per thousand keywords

Keyword grouping software is almost always priced per thousand keywords. It is the wrong unit. You do not consume keywords, you consume clusters you can actually brief an article from — and the number of those varies enormously between tools processing the identical list.

Work it through on 2,000 keywords. KeyClusters’ own blog puts its rate at $4.97 per 1,000 keywords, so that run costs $9.94. If you want search volume attached, that is extra and metered separately — SE Ranking, linked above, charges “$0.005 per query” for the volume lookup, which on 2,000 terms adds another $10, more than doubling the bill.

Now divide by what you got. Say the run returns 240 groups and 150 of them are coherent enough to hand to a writer. Your real rate is roughly seven cents per usable cluster. A tool that costs twice as much per thousand but returns 200 usable groups from the same list is the cheaper tool. This is the number to put in your scorecard, and almost nobody publishes it, because it can only be calculated after the output exists.

One more reason to do your own arithmetic: the published prices do not agree with each other. On the same day I checked, KeyClusters quoted $4.97 per 1,000 on its own blog while another page ranking for the same term listed the same tool at “$9 per 1,000 keywords” — 81% higher. Both pages rank. Both look authoritative. Price-check at the vendor’s own site before you build a business case, and remember that the output of this whole exercise is a set of content briefs, so the cost that matters is cost per brief you can commission.

Run the same list twice

Keyword grouping software that reads live results re-fetches them every time it runs. That means its output has a shelf life, and the vendor will not tell you how long.

Test it directly. Keep your answer key, run it through the same tool a week after the first pass, and score the two outputs against each other using the same pair-counting method. You are not measuring accuracy this time — you are measuring stability. A tool that returns substantially different groups from identical input is one you cannot build a content calendar on, however well it scored on day one.

Some drift is honest. Google states plainly that “positions in Google Search results aren’t static or fixed in place”, so any method that reads live results inherits that movement. The question is proportion. In my experience, a week-on-week agreement above 0.90 on a stable topic is normal; a sharp drop usually means the threshold is set near a cliff edge, where small ranking movements tip keywords in and out of groups. Loosening or tightening the dial by one step often fixes it.

Set your re-clustering cadence from this number rather than from habit. If your groups are stable across a month, re-running weekly is wasted budget. If they move every week, your topic is volatile and your content plan needs to be shorter-horizon than you thought.

What keyword grouping software will not tell you

Every tool described here assumes a stable page of ten blue links. That assumption is weakening.

Google’s AI Mode, in its own description, “uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf” — Google’s description of AI Mode is unambiguous about this. One prompt becomes several underlying searches. A grouping tool scoring overlap between the two terms a user might have typed is measuring something a step removed from what the system actually ran.

The practical consequence is not that grouping stops working. It is that the groups get larger than your tool suggests. If a single prompt fans out into five subtopic queries, the page that earns the citation is the one covering all five — which is usually a merge of several of the tight groups your tool returned at a high threshold. I have written separately on how query fan-out works and what it does to content planning.

So run your evaluation at the tight setting to understand the structure, then deliberately check what merges at the loose setting. Those merged groups are your candidates for the deeper pages that build topical authority in AI search. The tight groups tell you what to write; the loose ones tell you what to write at length.

The scorecard: choosing in about an hour

Put the whole test on one sheet and fill it in for each piece of keyword grouping software you are considering.

What you are measuringHow to get itPass mark
MethodAsk the vendor: word similarity or live search results?Must be stated plainly
ContaminationPairs the tool merged that your key separates≥ 0.80
SplittingPairs your key merges that the tool separated≥ 0.80
Threshold controlCan you change the overlap setting?Yes, and documented
StabilitySame list re-run a week later, scored against the first run≥ 0.90
Cost per usable clusterTotal run cost ÷ groups you would actually briefCompare, don’t threshold
ExportCan you get the groups out as CSV with your metrics attached?Yes

The first four take an hour once your answer key exists. Stability needs a week of waiting but no work. Cost falls out of the run you have already done.

A tool that refuses to tell you its method, or hides the threshold behind an “accuracy” slider with no numbers, has failed before you score it — not because it is necessarily bad, but because you cannot reason about its output. For the shortlist to run this against, I have compared the main options by the method each one uses in my keyword clustering tools roundup.

Frequently asked questions

Is keyword grouping the same as keyword clustering?

In practice the terms are used interchangeably, and the SERPs for them are genuinely different, so vendors optimise for whichever one they chose. The distinction worth caring about is not the word but the method: does the tool compare the keywords themselves, or the live results they return?

Can free keyword grouping software do the job?

For lists under a few thousand keywords, often yes. Free tools are usually word-similarity based, which is adequate when your keywords are already topically tight and inadequate when they are not. Run your answer key through the free option first — if it clears your pass mark, you have saved the subscription.

How many keywords should I group at once?

For testing, 100 to 150 is enough to be meaningful and small enough to hand-group. For production, the limit is whatever your tool charges for, but larger batches do not improve grouping quality: the tool compares pairs, so a 10,000-keyword run is not smarter than ten runs of 1,000 covering the same ground.

Does keyword grouping prevent cannibalisation?

It prevents the planning mistake that causes it. Grouping tells you which keywords one page should serve, so you commission one page rather than three. It does nothing about pages you have already published — that needs an audit of what currently ranks for what.

How often should I re-cluster?

Measure rather than guess. Re-run your list a week apart and score the agreement. Stable topics can go months; volatile ones need monthly runs. The number from your own stability test beats any vendor’s recommended cadence.