Search Results Clustering

Features

• Hosted app: connected components clustering with shared URLs slider (1-10, default 4)
• CLI script: three strategies (connected, cliques, core) or all at once
• Consolidation scoring 0-100 with recommendation bands
• Core strategy connectivity threshold (default 0.7, CLI only)
• Overlap detection and scoring penalty across clusters (CLI only)
• Clusters named after the shortest query; unclustered keywords kept as NO_CLUSTER

How to use

1. 1 Export SERP data as query and URL pairs (e.g. ValueSERP batch CSVs)
2. 2 Hosted app: upload CSVs; first column is read as the query, second as the URL
3. 3 CLI script: place Batch_Results_*.csv files (with search.q and result.organic_results.link columns) in the valueserp_exports folder
4. 4 Set the minimum shared URLs threshold (default 4)
5. 5 Run clustering and review clusters sorted by consolidation score
6. 6 Download the results CSV

Frequently asked questions

What exact input format is required?

The Python script reads CSVs named Batch_Results_*.csv from a valueserp_exports folder and requires two columns with the exact ValueSERP batch export headers: 'search.q' and 'result.organic_results.link'. The hosted Streamlit app is more forgiving: it takes whatever files you upload and uses the first column as the query and the second as the URL, ignoring header names.

Are all three clustering strategies available in the hosted app?

No. The Streamlit app implements connected components only, with a shared URLs slider (1 to 10, default 4). The cliques and core strategies, the CLUSTERING_STRATEGY setting and the overlap penalty exist only in the downloadable Python script.

Why does the same keyword appear in several rows of the script output?

With CLUSTERING_STRATEGY = 'all' (the default) the script runs all three strategies and outputs every cluster from each, so one keyword can appear up to three times under different cluster_type values. Set the strategy to 'connected', 'cliques' or 'core' if you want one row per keyword.

How is the 0-100 consolidation score calculated?

Base score of up to 40 from average shared URLs (average x 4, capped), plus connectivity x 30 (the share of keyword pairs in the cluster that actually meet the URL threshold), plus a size bonus of 5 points per keyword beyond two (capped at 20), minus an overlap penalty of 5 per query that sits in multiple clusters (capped at 10, script only). 80+ is a strong consolidation candidate, under 20 means keep separate.

How are clusters named and what does NO_CLUSTER mean?

Each cluster takes the shortest query in it as its name. Queries that share fewer than the threshold number of URLs with every other query are still included in the output, assigned to NO_CLUSTER with a score of 0 and the recommendation 'Keep separate', so the output always contains your full keyword list.