Problem
Who a site really competes with in search differs by topic, and a competitor list made by hand cannot be checked or rebuilt.
Value
A team sees its real search competitors per topic from data instead of assumptions, and every link in the map carries its source, so the result can be rebuilt.
Written from this project
Role
Method design, data model, evaluation design and the architecture decisions. Agents wrote most of the code.
- Defined the steps from ranking facts to site similarity and on to a sparse map of neighbouring sites, each step with its version and provenance.
- Designed a blind evaluation in which labelled site pairs test whether the neighbours make sense, against a simple baseline.
- Recorded the design in eight architecture decision records before the code grew.
Limits
- Built on data from paid SEO tools. Only the method and aggregate results are described; the files are not published.
- Paused since July 2026. The method moved into the Searchint work.
Sources
- Eight architecture decision records
- Private research repository