Building a competitor landscape corpus
The output: a market's competitor landscapes as one corpus, where every membership is discovery topology, not similarity, and the next seeds are chosen by the same rule every round.
- Updated
- ·
Inputs
-
The seed domains of a market: the sites whose competitors are wanted.
-
One competitor landscape per seed from a competitive-research tool's export (Semrush): the domains the tool discovers competing with the seed, with the database and the snapshot date. How the export is obtained is outside the method.
-
A list of already validated domains, so a known seed does not come back as a candidate.
-
A round budget: how many new landscapes one round exports.
Procedure
Start from the seed landscapes
The first round exports each seed's landscape. Every landscape is stored as it came: seed, discovered domains, database, snapshot date and whether the export was cut off.
Expand the frontier
The frontier is the set of discovered domains that are not yet seeds. Unvalidated ones stay outside it, and the next round exports the landscapes of those chosen from the frontier.
Keep the exclusive frontier apart
A domain that appears in only one landscape is the cheapest way to a new corner of the market. The exclusive share of a parent landscape tells how much new ground lies behind it.
Choose the next seeds by the same rule
A candidate is scored with versioned weights: its rank in the parent landscape, the landscape's row count, the parent's exclusive share, the parent's low overlap with others and the parent's sparsity. The same corpus and weights always give the same order, and the pick is spread across parents so one large landscape does not fill the round.
Record truncation and failures
An export that was cut off is marked, and its landscape is never presented as complete. Failed exports stay in the history so the same candidate does not loop forever.
Read membership as discovery, not proximity
A domain in many landscapes is well connected in discovery. It is not similar, not relevant and not in demand. The corpus stays a contract of its own beside the similarity graph: no combined weight between the two is defined.
Outputs and how to read them
-
The corpus: the landscapes with their memberships, snapshot dates and databases, every file checked against its schema.
-
A summary: counts of landscapes, domains and memberships, the frontier and the exclusive frontier, the failed exports.
-
The next round's seed list with its score components, so the choice can be checked.
Failure modes
-
Membership is read as similarity, and broad publishers look like everyone's competitor.
-
Landscapes from different snapshot dates mix in one corpus without a mark.
-
The frontier is expanded from the largest landscapes, and the market's edges never come in.
-
A cut-off export is presented as a complete competitor list.
-
The expansion never ends: the budget and the exclusive-share threshold are missing.
Where it is used
In the research layer for a Finnish market's search data: the corpus feeds diversified seed selection, and the similarity graph is evaluated apart from it.