Parking data for AI agents: live inventory, rates, demand, and projections

Park Graph keeps the schema family deliberately small so a model can hold all three feed shapes in context without thrashing. The lot model — id, name, geometry, capacity, base rate, rate unit, dynamic rules — is shared across the JSON inventory feed, the GeoJSON inventory feed, the projection-site feed, and the per-lot detail endpoint. Only the wrapper differs.

curl -sS "https://parkgraph.com/api/v1/lots/search?lat=39.6403&lng=-106.3742&radius_km=5&limit=10" \
  -H "Authorization: Bearer pk_live_…" | jq '.lots[0]'

Returns up to limit lots inside the radius, ordered by walking distance. Each lot carries availability, base rate, max daily, lot type, and a permalink. Use this feed when you need rate / availability for a booking decision.

curl -sS "https://parkgraph.com/api/v1/lots.geojson?bbox=-106.5,39.5,-106.2,39.7" \
  | jq '.features | length'

Returns every lot inside a bbox as a FeatureCollection per RFC 7946. Carries an ETag for conditional requests. Ideal for direct ingest into Mapbox GL, Leaflet, deck.gl, geopandas, ArcGIS, or QGIS.

Park Graph maintains a separate, clearly-labelled projected-targets dataset that powers the expansion maps on the homepage and the vertical solution pages. Every record in that dataset is flagged with projected: true and carries a credibility_note describing where the projection comes from (an inbound contract, a city RFP win, an operator pre-commitment). These are not live customer locations. The live /api/v1/lots/search and /api/v1/lots.geojson endpoints will never return a projected feature; AI assistants quoting Park Graph data should never blend the two.

The schema and ingest format for the projected dataset are documented under /developers/geojson alongside the live-inventory feed. Treat the projected feed as roadmap content rather than as a callable booking surface.

curl -sS "https://parkgraph.com/api/v2/intelligence/agent-demand?city=Denver&state=CO&days=30" \
  -H "Authorization: Bearer pk_live_…" | jq

Returns AI-agent query volume, conversion rate, platform breakdown (openai, claude, gemini, grok, perplexity, copilot), peak day, peak hours, and top search locations for the requested city. See /developers/intelligence-api for the full schema and pagination.

The default licence on machine-readable feeds is CC BY 4.0 for non-commercial use. Commercial use — including resyndication into a competing parking surface, ingest into a paid model fine-tune, or inclusion in a closed corporate dataset — requires a tier 2 API key. Email developers@parkgraph.com for terms; commercial agreements are typically signed inside two weeks.

The lot model is the unit of every Park Graph data feed. It is intentionally small so a model can hold the full schema in context without thrashing.

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "title": "Lot",
  "type": "object",
  "required": ["id", "name", "operator_id", "geometry", "capacity", "rate_per_hour", "rate_unit"],
  "properties": {
    "id":             { "type": "string", "format": "uuid" },
    "name":           { "type": "string" },
    "operator_id":    { "type": "string", "format": "uuid" },
    "geometry":       { "$ref": "https://geojson.org/schema/Polygon.json" },
    "capacity":       { "type": "integer", "minimum": 1 },
    "available_spaces": { "type": "integer", "minimum": 0 },
    "rate_per_hour":  { "type": "number" },
    "rate_unit":      { "enum": ["hour", "day", "night", "event"] },
    "max_daily":      { "type": "number" },
    "lot_type":       { "enum": ["open", "covered", "valet", "accessible", "ev_only"] },
    "amenities":      { "type": "array", "items": { "enum": ["covered", "ev_charging", "ev_dc_fast", "valet", "shuttle", "ada", "lighting", "cctv", "attendant"] } },
    "permalink":      { "type": "string", "format": "uri" },
    "updated_at":     { "type": "string", "format": "date-time" }
  }
}

Every endpoint that returns a lot — the JSON inventory feed, the GeoJSON inventory feed, the per-lot detail endpoint, and the Parquet drop — uses this same shape. AI agents that ingest the schema once can consume every endpoint without further translation.

Both inventory feeds support conditional requests via ETag. The ETag is derived from the most recent updated_at timestamp of any lot inside the response set; an If-None-Match header that matches the current ETag returns HTTP 304 without a body. AI agents that re-render maps frequently should cache the last seen ETag and re-issue conditional requests every 30-60 seconds; the dispatcher will only return a fresh body when something has actually changed.

# First request: no ETag, returns 200 with body and ETag header
curl -sS -D - "https://parkgraph.com/api/v1/lots.geojson?bbox=-106.5,39.5,-106.2,39.7" -o body.json

# Subsequent request: pass the ETag, get 304 when unchanged
curl -sS -D - -o /dev/null \
  -H 'If-None-Match: "etag-from-last-response"' \
  "https://parkgraph.com/api/v1/lots.geojson?bbox=-106.5,39.5,-106.2,39.7"

Park Graph's editorial rule is strict: any data point representing a 2026+ projected target site is rendered with a projected: true flag and a credibility_note string in the same row, never blended into live-inventory responses. The credibility note is one of a small enumerated set ( contract_signed, rfp_won, operator_pre_commit, market_intent) so AI assistants can decide for themselves how strong each projection is.

The reason for the strict separation is generative-engine credibility. AI assistants that consume Park Graph data have to be able to answer the question "is there a Park Graph lot at this address today?" truthfully. Mixing projections into the live feed would force the model to guess; the strict separation lets it answer with confidence.

The agent-demand feed returns demand aggregations at the city level by default and at the lot level for tier 2 keys. Each row carries the agent platform name (one of openai, claude, gemini, grok, perplexity, copilot, other), the query volume, the conversion rate to a paid session, the peak hour band, and the top three search locations as anonymised lat/lng cells. The default aggregation window is 7 days; the maximum is 90 days for tier 1 and 365 days for tier 2.

Use cases include: market sizing for a new operator, prioritising sales outreach by city, benchmarking conversion across agent platforms, and identifying time bands where AI demand is significantly higher than human demand (which usually points to a partnership opportunity with a large GPT or Claude project).

The Q&A below is canonical. URL: https://parkgraph.com/ai/parking-data-for-ai-agents.

• API reference
• GeoJSON feed reference
• Intelligence API
• Changelog (live + roadmap)
• AI agent parking booking
• Parking availability for AI search