How Alien Road Uses AI for Keyword Research and Clustering

A site that targets “running shoes for flat feet”, “best shoes for flat feet running” and “flat feet running trainers” with three separate pages often ends up with three weak pages competing for the same searches. Clustering finds that these phrases share one intent and should be answered by one strong page.

Manual clustering works for a few hundred keywords. For an e-commerce site or a large service business with tens of thousands of queries, it takes days, and results vary between the people doing it.

There are three common approaches, and they can be combined.

• SERP overlap: keywords are grouped when Google shows many of the same URLs for them. This reflects how Google itself sees intent. Tools such as Semrush’s Keyword Strategy Builder and several dedicated clustering tools use variants of this method.
• Embeddings: each keyword is turned into a numerical vector that captures meaning, using models from OpenAI, Google, Cohere or open-source libraries like Sentence Transformers. Similar vectors are grouped with algorithms such as k-means or HDBSCAN.
• Language model labelling: a model reads groups of keywords and assigns an intent label (informational, commercial, transactional, navigational) and a topic name.

SERP overlap is the most reliable signal of intent but costs more, because it needs search results for every keyword. Embeddings are cheap and fast but can group phrases that look similar and mean different things, such as “apple storage” for fruit and for iCloud.

Quality depends more on the steps around the AI than on the model itself.

• Collect: export keywords from Search Console, Keyword Planner and a third-party tool, plus questions from support and sales.
• Clean: remove duplicates, misspellings with no volume, and irrelevant terms such as competitor brands you will not target.
• Enrich: add search volume, current ranking URL and position where available.
• Cluster: run embeddings for a first grouping, then confirm close calls with SERP overlap.
• Label: use a language model to name each cluster and assign intent.
• Review: a strategist checks clusters, splits mixed ones and merges duplicates.
• Map: assign each cluster to an existing page or a new page in the plan.

Automated clustering makes predictable errors. A short checklist catches most of them.

• Mixed intent: a cluster containing both “how to fix a leaking tap” and “plumber near me” should be split; one is a guide, the other a service page.
• Ambiguous terms: check clusters with words that have several meanings.
• Over-broad clusters: a cluster with hundreds of keywords may need sub-topics.
• Brand terms: keep your own brand and competitor terms in separate groups.
• Language and location: for multilingual sites, cluster each language separately, since embeddings across languages can group translations together.

AI Overviews, AI Mode and assistants such as ChatGPT and Perplexity handle longer, conversational questions. Many of those queries have little or no volume in keyword tools, yet they reveal what buyers want to know. Adding real customer questions from support, sales calls and forums to the keyword list helps clusters reflect this.

Clusters also make it easier to plan answer content: one page that covers a topic thoroughly, with clear sections for the main questions, is more useful for both search results and AI answers than several thin pages.

AI clustering cannot tell you which topics are worth pursuing for your business; that is a strategy decision. It also depends on input data, so a keyword list missing whole product lines will produce a plan missing them too. And search behaviour changes, so clusters need refreshing, typically every few months for active sites.

Keyword clusters are a snapshot. New products, seasonal demand and changes in how Google groups results all shift them over time. For most active sites, a quarterly refresh using fresh Search Console data is enough, with a full rebuild once a year. Compare the new clusters with the old ones and look for three changes: clusters that have grown, new clusters with no mapped page, and clusters where the ranking URL has switched. Each points to a specific content task.