Keyword Clustering Tool

Paste your keyword list or a Search Console CSV and group it by topic and intent with AI that runs in your browser. Download the CSV. Free, no sign-up.

./keyword-clustering

If you paste a CSV with a header, the keyword/query column is used and, if present, the volume or impressions one.

The first time, the AI model (about 135 MB) is downloaded from Hugging Face and kept in the browser cache. On mobile data, better wait for wifi.

Keywords are processed with AI inside your browser; they are never sent to any server.

What this tool does

You paste a keyword list — or the CSV you export from Search Console, Ahrefs or Semrush — and the tool splits it into groups of searches that mean the same thing, or nearly so. «how much does a website cost», «website price for small business» and «how much do developers charge for a website» end up together even though they barely share words, and apart from «calculate days between dates».

For each group you see how many keywords it holds, the total volume or impressions if the CSV includes them, and a name: the keyword with the most volume or, without volume, the most representative one. The result downloads as a CSV with two columns, keyword and group, ready for a spreadsheet.

Why cluster keywords

A list of 500 queries does not tell you how many pages you need. Grouping it does:

  • One page per group. When several keywords ask for the same thing, splitting them across pages makes those pages compete. Grouping before writing avoids that cannibalization.
  • Prioritize by real demand. A group’s total volume or impressions says more than any single keyword.
  • Find gaps. A large group of Search Console queries with no page serving it is the next piece to write.

How to use it

  1. Paste your keywords, one per line, or a CSV with its header. You can also load the file with Open CSV. Recognized columns are «Keyword», «Top queries» / «Consultas principales» and, for volume, «Volume», «Volumen», «Impressions» or «Impresiones». Duplicates are dropped automatically.
  2. Click Cluster keywords. The first time, the model is downloaded; you will see the progress.
  3. Move the Minimum similarity slider until the groups make sense for your case. The change is instant: the model does not run again.
  4. Download the CSV.

Keywords not similar enough to any other end up under Ungrouped. That is not an error: they are usually topics that appear only once in your list.

Choosing the threshold

There is no single right number for every list. As a guide:

  • Further left (broad groups): joins everything about the same topic. Useful to see the big content blocks of a site.
  • Further right (strict groups): only joins near-identical variants of the same search. Useful to decide which keywords a single page should target.

If a group mixes two intents you would handle separately — say «price» and «how to» — raise the threshold a little.

Your keywords never leave your device

Clustering is done by an AI model that runs inside your browser. The list and the CSV are not uploaded to any server, there is no account and no daily limit, and you can work with a client’s Search Console queries without sharing them with a third party. Once the model is downloaded, the tool works even offline.

Requirements and download size

  • The model is paraphrase-multilingual-MiniLM-L12-v2, trained on more than 50 languages, in its quantized version. A multilingual model was chosen because small English-only models cluster other languages poorly.
  • The first time it downloads about 135 MB from Hugging Face (the 118 MB model and its 17 MB tokenizer). They stay in the browser cache and are not downloaded again. On mobile data, better wait for wifi.
  • It runs on the CPU in the background, so the page does not freeze. A hundred keywords take a few seconds; 2,000 may take around a minute on a modest machine.
  • The limit is 2,000 keywords per batch. If you paste more, the tool warns you and clusters the first 2,000.

What it does not do

It does not look up search volume, difficulty or the SERP: it only works with what you paste. Nor does it check which pages rank for each keyword, which is how some paid tools cluster; here the criterion is the meaning of the search. For most lists the two agree, but if two similar keywords have very different Google results, the model cannot know.

Frequently asked questions

How does it decide which keywords belong together?

An AI model turns each keyword into a vector that represents its meaning, so «website price» and «how much does a site cost» end up close even without shared words. Then hierarchical clustering joins the most similar ones while their similarity stays above the threshold you set with the slider.

Are my keywords sent to a server?

No. The model is downloaded once from Hugging Face and then runs inside your browser. The list, the CSV and the result never leave your device, so you can cluster a client's Search Console queries without sharing them with anyone.

How big is the download and how many keywords does it take?

The first time it downloads about 135 MB (the model and its tokenizer), which stays in the browser cache. It takes up to 2,000 keywords per batch; if you paste more, it warns you and clusters the first 2,000.

Does it work with non-English keywords?

Yes. It uses paraphrase-multilingual-MiniLM-L12-v2, a model trained on more than 50 languages, because small English-only models cluster other languages poorly. You can mix languages in the same list.

Which threshold should I use?

Start with the default and move it while watching the table: to the left groups get broader (by topic), to the right stricter (near-synonyms, useful to decide which keywords a single page should target). The change is instant; the model does not run again.