Keyword research dies in spreadsheets. Somewhere around row three thousand your eyes glaze, every phrase looks the same, and the clustering you promised yourself becomes a color-coding exercise you abandon by Friday. Claude Code keyword research replaces that grind with one instruction over a folder of exports, and what comes back is a clustered, deduplicated, intent-labeled plan instead of a bigger spreadsheet.
Nothing changes about your data sources. SEMrush, Search Console, wherever your numbers come from today.
What changes is everything after the export.
Claude Code keyword research runs in three steps: export keyword data from your tools into one folder, tell Claude Code to merge, clean, and cluster it by search intent, then have it map every cluster to an existing URL or mark it as a new page. Volumes stay tool-sourced. Claude Code does the sorting, clustering, and mapping.
Why Move Keyword Research Into Claude Code
The chat version works, and the clustering prompt in our Claude prompts for SEO collection is exactly that. It hits a wall at scale. Big exports exceed what a chat comfortably holds, and next month you paste it all again.
Claude Code reads the files directly, which is the same advantage that moved our audits into the terminal. No paste ceiling, no re-uploading, and the output lands as a CSV you can open, sort, and hand to a writer.
The repeatability is the part nobody advertises. Same folder, same instruction, every month, and the results diff cleanly compared to last month.
The handoff improves as a side effect. clusters.csv opens in any spreadsheet, the writer gets a clean sheet instead of a chat transcript, and nobody asks which version of the clustering is current.
Repeatability is also what makes the habit stick. Hand-built clustering gets skipped the first busy month, and skipped work never compounds. A one-instruction run survives busy months.
Claude Code Keyword Research Workflow
One folder per market. Exports go in, a plan comes out.
Read every CSV in this folder. Then:
1. Merge them and drop exact duplicates
2. Remove branded terms, the brand list is in brand.txt
3. Cluster what remains by search intent, one cluster per topic a single page could own
4. Map each cluster to a URL from sitemap.txt, or mark it NEW PAGE
5. Keep volume and difficulty exactly as they appear in the source files
Write the result to clusters.csv with columns:
cluster, intent, top keyword, total volume, difficulty range, target URL, action
Two support files do a lot of quiet work here. brand.txt holds every brand spelling you want excluded, and sitemap.txt is just your URL list, so mapping happens against reality instead of memory.
An honest scar: our first multi-market run merged UK and US exports and double counted almost everything. Since then, one folder per market, no exceptions.
Drop a Search Console queries export into the same folder and the mapping step gets sharper. Clusters that already earn impressions are demanded if you have half-captured, so we have the instruction to add a priority flag wherever GSC and the tool export overlap.
What The First Run Gets Wrong
Expect two misses. Intent labels wobble on short ambiguous heads, a query like “roof cost” can be researched or purchased, and Claude picks one without telling you it guessed.
And clusters come back too granular, five clusters where one page should own all five. The fix for both lives in the instruction, not in the output. Add a merge rule, name the ambiguous cases, rerun. Next month inherits every correction, which is the quiet advantage of instruction-driven work.
| Step | What happens | Output |
| Merge and clean | Duplicates and branded terms removed | keywords_clean.csv |
| Cluster | Intent groups a single page can own | clusters.csv |
| Map | Each cluster matched to a URL or flagged new | action column in clusters.csv |
| Diff | Compared against last month’s run | changes.md |
The Monthly Diff, Where This Pays Off
The first run gives you a plan. The runs after that give you a moving picture, and the moving picture is worth more.
changes.md answers three questions each month. Which clusters are new, which grew, which shrank. A new cluster is demand forming before your competitors have a page for it. A shrinking one saves you from writing content into a fading topic.
We caught our first rising cluster this way, a question pattern that had barely existed in the previous export. The post written that week ranked before the topic got crowded, which is the entire argument for running research monthly instead of once per engagement.
Seasonal wobble shows up too. Learn to read which shrinkage is decay and which is December.
Can Claude Code Replace ahrefs or semrush?
No, and be suspicious of anyone who says otherwise. Claude Code has no search index and no volume data of its own.
Every number it should touch must arrive in your exports, and the instruction above tells it to keep those numbers untouched. Its job is the sorting and mapping layer that your keyword tool was never good at.
Ask it to estimate a volume and it will produce one. That number is fiction. This rule has no exceptions.
Reading the Clusters like an SEO, Not a Spreadsheet
The CSV is not the deliverable. The judgment pass is. We look for four signals before a cluster earns a slot on the calendar.
- Real search intent you can name in one sentence
- A clear gap, the ranking pages answer the query badly or not at all
- Fit, you can write it with first-hand experience rather than research theater
- A conversion path, some clusters bring traffic, the best ones bring clients
Score what passes before you schedule anything. We rank by volume multiplied by fit, then cut the list in half. The discipline is refusing to write everything, because forty average posts lose to twenty posts written from real experience.
Clusters that pass become briefs, and the brief prompt in the prompts collection picks up exactly where clusters.csv leaves off. The strategy layer above all of it, how many clusters, which lanes, what cadence, is the subject of how to use Claude for SEO.
Run it once on last month’s exports and compare against the clusters you built by hand. Where they disagree is where you will learn something, sometimes about the tool, sometimes about the hand-built version.
The cost side, since people ask. Setting up the folder, the brand list, and the first instruction took us about an hour. The monthly rerun takes twenty minutes, most of it spent reading changes.md over coffee. Against the two days a careful manual clustering used to take, the math is not close.
FAQ
How big an export can this handle?
Tens of thousands of rows is routine, since it processes files in chunks. Keep markets in separate folders and size stops being a concern.
Does this work with search console data too?
Yes, and the combination is the point. GSC shows queries you already surface for, and the mapping step marries them to tool exports cleanly.
How often should the clustering run?
Monthly fits most sites. The diff against last month is where trends show up, new clusters appearing is demand you can catch early.
SEO and digital marketing specialist. I also write about jobs abroad, relocating overseas, and learning languages. MBA in Marketing.