Site Selection and Cannibalization Analysis

Existing locations + Target area -> Spatial indexing -> Enrich with demographics/POIs -> Score/Rank -> Filter top candidates -> Save

Step 1: Load Data

Load two datasets with native.gettablebyname:

• Existing locations (current stores/facilities)
• Target area (e.g. city boundary, district polygons, or a grid covering the study area)

Success: Both tables loaded with geometry columns and unique identifiers.

Step 2: Build Candidate Grid

Polyfill the target area into H3 or Quadbin cells using native.h3polyfill or native.quadbinpolyfill. Each cell is a candidate micro-location.

Success: A contiguous grid of cells covering the study area.

Step 3: Enrich Candidates

Attach demand signals to each cell — population, income, foot traffic, POI density — using native.h3enrich, native.joinv2, or the Data Observatory.

Success: Each grid cell has numeric columns representing demand/suitability factors.

Step 4: Filter by Proximity to Existing Locations

Use native.h3distance to compute hop distance from each candidate cell to the nearest existing location. Filter out cells that are too close (cannibalization risk) or too far (logistics cost).

• native.h3distance returns hop count, not physical distance. Convert using the approximate edge length for the resolution (e.g. H3 res 8 ~ 460m edge, so 3 hops ~ 1.4 km).

Success: Candidate cells are within a sensible distance band from existing locations.

Step 5: Score and Rank

Use the scoring pattern from trade-area-analysis:

1. Normalize each variable to [0,1] with native.normalize
2. Composite score via native.selectexpression with user-defined weights
3. Rank with native.orderby (descending) + native.limit (top N)

Success: A ranked shortlist of candidate cells with composite scores and contributing variables.

Step 6: Save

Use native.saveastable. The H3/Quadbin column is directly visualizable in CARTO Builder.

Success: Validated workflow ready to upload.

Pattern B: Cannibalization Analysis

Existing + Proposed locations -> Trade areas (isoline/buffer) -> Polyfill to grid -> Intersect/Join -> Measure overlap -> Save

Step 1: Load Data

Load existing locations and proposed locations (or a single table with a flag column distinguishing them).

Success: Both sets loaded with geometry and unique identifiers.

Step 2: Generate Trade Areas

Create catchment areas around both existing and proposed locations using native.isolines (realistic) or native.buffer (simple). Use the same parameters for both sets to ensure comparability.

Success: Every location has a catchment polygon with consistent parameters.

Step 3: Polyfill to Spatial Index

Convert all catchment polygons to H3 or Quadbin cells with native.h3polyfill. Preserve the location identifier and an is_proposed flag.

Success: One row per cell per location, with location ID and type flag.

Step 4: Find Overlap

Use native.joinv2 (inner join on the spatial index column) between existing-location cells and proposed-location cells. The result contains cells shared by at least one existing and one proposed location.

Success: Output contains only cells that fall in both an existing and a proposed catchment.

Step 5: Measure Impact

Use native.groupby to aggregate overlap:

• Per existing location: count of overlapping cells / total cells in that location's catchment = overlap percentage
• Enrich overlap cells with population or revenue to quantify shared demand

Use native.selectexpression to compute the overlap ratio.

Success: Each existing location has an overlap metric showing how much of its catchment is shared with proposed locations.

Step 6: Save

Use native.saveastable.

Success: Validated workflow with per-location cannibalization metrics.

Pattern C: Twin Areas / Similar Locations

Top-performing locations -> Trade areas -> Enrich -> Build similarity model -> Score all candidate areas -> Rank -> Save

Step 1: Identify Reference Locations

Load the full location dataset. Filter to top performers (e.g. top quartile by revenue) using native.wheresimplified or native.orderby + native.limit.

Success: A subset of high-performing locations isolated as the reference set.

Step 2: Generate and Enrich Trade Areas

Create isochrone or buffer trade areas around reference locations. Polyfill to H3/Quadbin. Enrich with demographics, POIs, and any relevant variables.

Success: Each reference location has a rich demographic profile.

Step 3: Build Twin Areas Model

Use native.buildtwinareasmodel (BUILD_TWIN_AREAS_MODEL) to create a PCA-based similarity model from the enriched reference locations.

• Input: enriched reference locations with numeric feature columns
• The model captures the multivariate "signature" of successful locations

Success: A model artifact that encodes the demographic profile of top performers.

Step 4: Find Similar Locations

Use native.findsimilarlocations (FIND_SIMILAR_LOCATIONS) to score all candidate areas against the twin-areas model.

• Input: candidate areas enriched with the same variables used to build the model
• Output: similarity score per candidate

Success: Every candidate area has a similarity score relative to the reference set.

Step 5: Rank and Save

Rank by similarity score descending. Save top candidates.

Success: A ranked list of areas most similar to top-performing locations.

Commercial Hotspots Variant

For demand-driven site selection (e.g. "where is unmet demand highest?"), use native.commercialhotspots:

1. Build an H3 grid over the study area
2. Enrich with the target demand variable (e.g. population aged 15-34)
3. Run native.commercialhotspots with variablecolumns and weights
4. Filter results by significance (p_value < 0.05)
5. Optionally filter by native.h3distance from existing locations to focus on underserved areas

Note: variablecolumns uses Python-style list syntax (['col1', 'col2']), and weights is comma-separated — see the trade-area-analysis gotchas for details.

Gotchas

• native.commercialhotspots requires the Retail module of the Analytics Toolbox. Validate with --connection to confirm availability.
• Twin Areas and Similar Locations use PCA internally — results are sensitive to variable selection and scaling. Include only relevant, non-redundant variables. Normalize inputs if scales differ widely.
• Cannibalization overlap depends heavily on trade area definition (buffer radius, isoline time). Small changes in parameters can flip results. Document the chosen parameters and rationale.
• native.h3distance returns hop count, not physical distance. Multiply by the approximate cell edge length for the resolution to get a rough metric distance (e.g. res 8 ~ 460m, res 9 ~ 174m per hop).
• When comparing across regions of different sizes, normalize demographics to per-capita or per-area values to avoid size bias (e.g. population density instead of total population).
• The "best" location depends entirely on the criteria and weights chosen — there is no objectively correct answer. Always document assumptions and let the user adjust weights.
• For the twin-areas model, use the same set of enrichment variables for both the reference locations and the candidates. Mismatched variables will cause the model to fail or produce meaningless scores.

Reference Templates

Academy Tutorials

Common Variations