The Keyword Clustering Prompt

The Prompt

Attach your keyword CSV and paste this. It expects four columns: keyword, monthly search volume, competition, and top-of-page bid. It works with fewer, it just has less to reason about.

Attached is a keyword export (keyword, monthly search volume, competition, top-of-page bid).

Work through these steps in order and show me the output of each.

STEP 1: DEDUPE
Collapse exact duplicates and near-identical variants (plural/singular, word order,
spacing). Keep the highest volume of the set. Tell me how many rows I started with
and how many unique keywords remain.

STEP 2: CLUSTER BY GOAL, NOT BY WORDS
Group keywords so that every keyword in a cluster could be satisfied by ONE page.
Cluster by what the searcher is trying to accomplish, not by shared words. Two
keywords sharing a word but wanting different outcomes belong in different clusters.
Two keywords with no words in common but the same goal belong together.

Create a separate cluster for misspellings and typos of the core term. Do not
discard them and do not merge them into the correct-spelling cluster.

STEP 3: LABEL INTENT
Tag every keyword with exactly one:
- transactional (ready to act now)
- commercial (comparing, evaluating, close to acting)
- informational (learning, not buying)
- navigational (looking for a specific named brand or product)

STEP 4: ASSIGN ONE PAGE PER CLUSTER
Give each cluster a single page slug. If two clusters would need the same page,
merge them. If one cluster needs two pages, split it.

STEP 5: TELL ME WHAT NOT TO BUILD
Flag any cluster where the searchers are NOT my audience, even if the volume is
large. Mark those "DO NOT BUILD" and say why in one line. Do not soften this.
High volume with the wrong intent is a trap.

STEP 6: OUTPUT
A table sorted by total volume, descending:
cluster | # keywords | total monthly volume | dominant intent | avg top-of-page bid | page slug | verdict

Then, separately:
- the intent split: keyword count AND total volume for each of the four intents
- any cluster with mixed intent, flagged as needing a split
- a leftover bucket for anything you could not confidently classify, with the count

Do not invent volumes. Use only the numbers in the file. If a keyword has no
volume, say so rather than estimating.

The Three Instructions Doing the Real Work

Most clustering prompts are a single sentence: group these keywords by topic. That produces a tidy table and no judgment. Three specific instructions above are what change the output from a list into a decision.

• Cluster by goal, not by words. Without this, the model groups on string similarity, which is the one thing a spreadsheet formula could already do. Goal-based grouping is the entire value of using a language model here.
• Keep misspellings separate. Without this, typo variants either get absorbed into the main cluster or dropped as noise. Either way you never see how much demand they represent.
• Tell me what not to build. This is the instruction no tool has. It asks for a verdict, not a grouping, and it is the only step that can save you from writing forty pages for an audience that was never going to convert.

What It Produced on a Real Export

We ran this against a Google Keyword Planner export from a background-removal SaaS: 32,632 rows, of which 16,608 were duplicate variants that the tool had grouped and re-listed. After deduplication, the top 2,000 unique keywords by volume went through the prompt in a single pass and came back as 16 clusters, every keyword tagged with an intent and assigned to a page.

Six smaller clusters sat below these ten: object and watermark removal, blur background, mobile app, use-case specific removal, signature and document photos, and a 90-keyword review bucket for terms that could not be confidently classified. A leftover bucket is not a failure of the method. Any honest clustering pass has one, and a tool that reports zero unclassified keywords has quietly forced some of them into the wrong group.

The Finding That Only Intent Reveals

Across those 2,000 keywords there were 880 commercial keywords and 882 informational keywords. Counted as keywords, the two groups are indistinguishable. Counted as demand, the commercial keywords represent 115.9 million monthly searches and the informational keywords represent 5.6 million.

That is the whole argument for clustering by intent rather than by volume or by topic. A keyword count treats those two buckets as equal priorities. The intent column shows that one of them is roughly twenty times the opportunity of the other, and it tells you which pages to build first.

The second finding was sharper. The single largest cluster by keyword count that was not core product demand, 370 keywords carrying 2.4 million monthly searches, turned out to be people searching for background images to download for photo editing. The product removes backgrounds. Same vocabulary, opposite goal, and enough volume to look like the biggest opportunity on the map. It was marked DO NOT BUILD, and that verdict was worth more than the other fifteen rows combined.

Where This Breaks

Two honest limits. Around 2,000 unique keywords is the practical ceiling for a single pass, so a 30,000-row export needs deduplication and a volume floor before it becomes tractable. And the prompt reasons about the numbers in your file, so it inherits whatever is wrong with them. Keyword Planner groups close variants and reports them at a shared volume, which is why one term can appear nine times at an identical figure. Deduplication in step one is not housekeeping. It is what stops the same demand being counted nine times.

Frequently Asked Questions