AI site selection with GIS: a practical guide to finding and ranking sites

Why site selection takes so long the traditional way

A typical screen for a new facility needs six to ten layers: parcels, zoning or land use, flood hazard, roads, utilities, competitors or existing facilities, and demographics. Each lives on a different portal, in a different format and projection. Parcel data alone can mean a county assessor download, a schema you have never seen and a field called something like LUC_CD that nobody documented.

Only after all of that does the actual analysis start, and the analysis is the easy part: a few buffers, an overlay, a join and a sort. NeuraCities moves the effort back where it belongs, onto deciding the criteria.

Step 1: write the criteria down

Good site selection starts with separating what is non-negotiable from what is merely preferred. For example, for a mid-size data center:

You do not need to fill in the data column yourself. That is what the Data Agent is for. But being explicit about which criteria are filters and which are preferences makes the result far easier to defend.

Step 2: ask the question

Type it the way you would brief a colleague. For example:

Find parcels in Houston of at least 10 acres that are within 1,500 m of a substation and 500 m of a fiber line, outside the FEMA 100-year floodplain. Rank them by zoning compatibility and distance to the nearest highway, and show the top 20.

NeuraCities responds with a checklist of the data and steps it will use before doing anything else. If it has read a criterion differently than you meant (for example, the 0.2% annual chance zone instead of the 1% zone), you can correct it right there.

Step 3: what the Data Agent pulls in

For a US site selection question, the Data Agent typically assembles:

• Parcels from the county or city open data portal, or from statewide parcel programs where they exist.
• Zoning and land use from the city's portal.
• Flood hazard zones from the FEMA National Flood Hazard Layer.
• Infrastructure such as substations and transmission lines from federal and local sources, and fiber routes where a city or state publishes them.
• Roads, rail and amenities from OpenStreetMap.
• Demographics such as population, income and households from the Census American Community Survey, by tract or block group.
• Competitors or existing facilities from business listings or open data.

Each dataset is probed for coverage in the study area before it is downloaded, and appears on the map as its own layer with its source attached. If a key layer does not exist publicly (fiber is often the gap), NeuraCities tells you, and you can upload your own.

Step 4: filter, score and rank

The GIS agent turns the criteria into a sequence of standard operations:

1. 1Buffer substations by 1,500 m and fiber by 500 m.
2. 2Keep parcels that intersect both buffers.
3. 3Remove parcels that intersect the 1% annual chance flood zone.
4. 4Filter by calculated parcel area.
5. 5Join zoning to each remaining parcel and measure the distance to the nearest highway.
6. 6Normalise each soft criterion and combine them in a weighted overlay using the weights you gave (or equal weights if you did not).
7. 7Sort by score and return the top candidates.

Because each step is a deterministic GIS operation, the result is repeatable: run it again on the same data and you get the same parcels. Change a weight and only the ranking changes. In one real Houston run of this kind of screen, the combined power and fiber proximity rule narrowed Houston's parcels down to 34 candidates, a list short enough to review by hand.

For demand-driven sites (retail, clinics, childcare), the analysis usually adds a catchment: population within a network service area around each candidate, computed on the street network and summed with an area-weighted join to the Census geography. That tells you how many people each site actually serves, not just how close it is to something.

Step 5: review the shortlist

The answer arrives as a working set of outputs:

• A map of the candidates, styled by score, over the constraint layers.
• A ranked table with parcel IDs, area, zoning, distances and score, exportable to Excel.
• Charts comparing the top sites on each criterion.
• A written summary with counts, the criteria as applied and the sources used.

neuracities.com

From here, keep asking. "Drop anything within 300 m of a school." "Show me the top five in 3D against the satellite basemap." "Which of these are owned by the same entity?" Each follow-up builds on the layers already loaded, so it is fast.

Step 6: turn it into a deliverable

• Map layouts: the export canvas builds print-ready pages with legends, titles and annotations, as PNG, JPG, SVG or PDF, and multi-page decks for a site package.
• Dashboards: turn the conversation into a long-form page with the maps and charts embedded, edit the text, and share it by link.
• GIS files: download the candidate layer as a Shapefile or CSV for your GIS team or CAD workflow.
• Live sharing: send a link to the conversation so a colleague or client can open the same maps and ask their own follow-ups, and leave comment pins on specific parcels.

Step 7: make it repeatable with workflows

Most teams run the same screen many times: the same criteria for a different metro, or the same metro with a different acreage. Save the prompt as a workflow and replace the parts that change with variables:

Find parcels in {{city}} of at least {{min_acres}} acres within {{power_distance}} of a substation, outside the FEMA 100-year floodplain, ranked by distance to the nearest highway.

When you run the workflow, each {{variable}} becomes an input field, including a location picker for the city. The analysis is identical every time, which is exactly what you want when comparing markets.

Site selection questions teams run in NeuraCities

Tips for a defensible analysis

• State distances and thresholds explicitly. "Near a highway" is a judgement; "within 2 km of an interstate on-ramp" is a criterion.
• Check the sources on the key layers. Open the layer details to see who published it and how current it is, especially parcels and zoning.
• Upload what you trust more. If your client has a better parcel file or a utility has shared its lines, upload it; your data takes priority.
• Keep the weights visible. Put them in the question so they are recorded in the conversation and anyone reviewing the result can see them.
• Treat it as a screen, then verify. A GIS screen narrows hundreds of parcels to a handful. Title, environmental and engineering due diligence on the finalists still matter.

Frequently asked questions