Measuring AI for keyword research and clustering means checking two things: whether the clusters the tool creates are correct enough to build pages on, and whether the content built from them brings more rankings, traffic and conversions per cluster. Speed matters too, but a fast tool that groups keywords wrongly only produces wrong pages faster.
AI keyword clustering groups related search terms so that one page can target a whole set of queries with the same intent. Tools do this in two main ways. SERP-based clustering compares the search results for each keyword and groups terms whose top results overlap. Semantic clustering uses language model embeddings to group terms by meaning. Keyword Insights, SE Ranking, Semrush’s Keyword Strategy Builder and many custom Python setups use one or both.
Why measure the tool at all?
Clustering decisions shape the site. If a tool puts “crm for small business” and “free crm” in one cluster, you may write one page where the results show two different intents. If it splits “crm software” and “crm system” into separate clusters, you may write two pages that compete with each other.
Both errors cost months. Measuring cluster quality before writing protects the content budget, and measuring results afterward tells you whether the method should be kept. It also gives the team a shared language: instead of arguing about whether a page is “good”, people can point to a cluster score and a ranking breadth number.
How do you check cluster quality?
Take a sample. For a list of 5,000 keywords, review 20 to 30 clusters by hand, chosen at random across sizes. For each cluster, search two or three of its keywords and look at the top results.
- Precision: do all keywords in the cluster share the same intent and similar results?
- Completeness: are obvious variants missing that ended up in other clusters?
- Intent label: is the cluster correctly marked as informational, commercial, transactional or navigational?
- Size sanity: are there giant clusters that mix topics, or hundreds of single-keyword clusters?
Score each cluster as correct, needs merge, needs split or wrong intent. Two reviewers working independently on the same sample is better than one, because it shows where intent is genuinely ambiguous rather than where one person made a quick call. A simple percentage of correct clusters gives you a quality number you can compare between tools or settings.
SERP-based or semantic: which should you trust?
SERP-based clustering reflects how Google currently groups intent, which makes it reliable for deciding what goes on one page. It costs more because each keyword needs a results check, and results differ by country.
Semantic clustering is cheaper and faster, and it handles large lists well. It can be wrong where meaning is close but intent differs, such as “apple watch bands” and “apple watch band sizes”. Many teams use semantic clustering to explore a large list, then SERP-based checks on the clusters they plan to write.
Which KPIs prove the method works?
Once pages are built, measure at cluster level rather than keyword level. Clusters are the unit the work was planned around, so they are the right unit to report.
Cluster-level KPIs
- Coverage: share of target clusters that have a live, indexed page
- Ranking breadth: number of keywords in the cluster where the page ranks in the top 10 or top 20
- Cluster traffic: organic clicks from Search Console for queries in the cluster
- Cannibalisation rate: clusters where two or more of your URLs swap positions for the same query
- Conversions: leads or sales from the cluster’s landing pages in GA4
- Time to first top-20 ranking after publishing
Ranking breadth is often the clearest sign of good clustering. A page built on a correct cluster tends to rank for many of the cluster’s terms, not only the main one. If a page ranks only for its head term after three months, check the cluster again before rewriting the content; the problem may be the grouping, not the writing.
How do you detect cannibalisation from bad clusters?
In Search Console, open the Performance report, filter by a query and switch to the Pages tab. If two or more URLs receive impressions for the same query over the same period, check whether they target the same intent.
SEO platforms also offer cannibalisation reports. Track the number of affected clusters monthly. If it rises after a clustering-led content push, the clusters were probably split too finely.
Building the dashboard
Keep keyword-to-cluster mapping in a sheet or database with columns for cluster ID, intent, target URL and publish date. Join this with Search Console query data through the Search Console API or a connector in Looker Studio or BigQuery.
The main view shows clusters as rows with coverage, ranking breadth, clicks and conversions as columns, sorted by value. A second view lists cannibalisation cases. A third tracks cluster quality scores from the manual review each quarter, so the team can see whether tool settings are improving.
How Alien Road applies AI keyword clustering
In a typical engagement, discovery means collecting keyword data from Search Console, SEO platforms and the client’s own site search, then removing branded noise and irrelevant terms. We run a semantic pass to explore the list and a SERP-based pass on the clusters that matter commercially.
Setup includes the manual quality review described above, a mapping sheet that links each cluster to one URL and a content brief per cluster. Measurement follows the cluster-level KPIs, reported monthly, with a quarterly review of cluster quality and cannibalisation. When results show that a cluster was wrong, we fix the mapping and the page, not just the report.
Where to start this week
Take your current keyword list and pick 20 clusters at random. Search two keywords from each and mark whether they belong together. The percentage you get is your baseline quality score.
Then open Search Console and check five of your most important queries for multiple ranking URLs. If you find cannibalisation, trace it back to how those pages were planned. It is the fastest way to see if your clustering method needs fixing.
Keep reading: SEO (Search Engine Optimization) · SEO