What is POI Data? Points of Interest Data, Sources

POI (point of interest) data is a dataset of the places people visit or search for: shops, restaurants, hospitals, schools, parks, and landmarks. Each record pairs a location with attributes such as a name, a category, an address, opening hours, a phone number, and a website. Retailers, insurers, mobility analysts, and map applications use point of interest data to find where things are and what sits nearby. The hard problem in POI data is identity: deciding which records describe the same place, and keeping that identity stable as places open, move, and close. The Chicago walkthrough below tests the first half. It looks for duplicate records in one release of Overture Maps, an open map dataset.

Key takeaways

• A POI record is a point plus attributes: name, category, address, contact details, hours, operating status, and often a confidence score.
• Overture Maps places holds 81,455,423 places from nine providers in its September 2026 release.
• In downtown Chicago, 91.0% of Overture's 52,917 places carry a website or a phone, and 14.3% have no category.
• In the same area, 58 pairs of places sit within 25 m of each other under near-identical names.

What is POI data?

A point of interest is a term from cartography. The OpenStreetMap wiki describes it as a feature shown with an icon at a single point, such as a post office, a shop, a pub, or a tourist attraction.

POI data turns those map icons into a table. Overture defines a place as "a concrete, physically identifiable, stationary destination in a publicly observable space." That definition leaves out administrative areas such as cities, movable things such as food carts, waypoints such as bus stops, and private residences.

In a geographic information system (GIS), a POI is a point feature with a name and a category. It is stored in a layer beside polygons such as parcels, census tracts, and building footprints. Apache POI is an unrelated Java library for reading and writing Microsoft Office files.

What a POI record contains

Attribute	Example fields
Identity	Stable ID, source record IDs, release version
Name	Primary name, alternate and translated names
Category	Primary category, category hierarchy, alternate categories
Location	Point geometry (longitude, latitude), sometimes a building polygon
Address	Street, locality, postcode, region, country
Contact	Phone numbers, websites, email, social profiles
Hours	Opening hours by day, holiday exceptions
Brand	Brand name and a link to a brand registry such as Wikidata
Status	Open, temporarily closed, or permanently closed
Quality	Confidence score, source datasets, last update time

Sources fill these fields differently. Foursquare's OS Places schema adds dates. date_created is when a place entered the database. date_refreshed is the last time a crawl, a user, or a human check refreshed any reference to it. date_closed is when it was marked closed. OpenStreetMap stores hours in an opening_hours tag, and Google returns hours with a business status. Commercial schemas add more: SafeGraph's places schema includes a Placekey identifier, brand IDs, and a polygon geometry for each place.

POI category taxonomies

Each provider keeps its own category list. Overture's Places Categories taxonomy has roughly 2,300 categories under 13 top-level groups. Each place's hierarchy array runs from the top level to its primary category, so a query can roll Cantonese restaurants up to Chinese restaurants, restaurants, and food and drink. A separate basic_category field holds about 280 labels, such as restaurant or museum. It draws on the basic-level categories that Rosch and colleagues described in Basic objects in natural categories (Cognitive Psychology, 1976). The basic level is the one people name first: "chair" before "furniture" or "kitchen chair".

Foursquare Open Source Places uses more than 1,000 categories in a hierarchy up to six levels deep. OpenStreetMap has no closed list: mappers combine tags such as amenity=cafe or shop=bakery. Google gives each place a primary type and a list of place types drawn from two fixed tables. Moving between them takes a crosswalk, a table pairing categories across lists. Overture maps each provider's categories into its own taxonomy on load, with exact lookups and reviewed model-derived matches.

POI data sources and providers

1. Overture Maps places. Free to download as GeoParquet, an open file format for map data, and updated monthly on the AWS Open Data Registry. The license is Community Data License Agreement (CDLA) Permissive 2.0 and Apache 2.0. It contains no OpenStreetMap data. The September 2026 release combines 58,783,121 records from Meta, 10,255,071 from BrightQuery, 6,135,466 from Microsoft, 4,138,835 from Foursquare, 1,809,219 from AllThePlaces, and smaller contributions from PinMeTo, DAC, Krick, and RenderSEO.
2. OpenStreetMap. Volunteers map POIs as tagged nodes and areas under the Open Database License, with coverage that varies by community.
3. Foursquare OS Places. More than 100 million POIs in over 200 countries and territories, released under Apache 2.0 in November 2024 and updated monthly.
4. Commercial POI data providers. Companies such as SafeGraph, Placer.ai, Dataplor, and Mapbox license curated places with verified brands, store polygons, and in some products foot traffic. SafeGraph sold its Patterns foot traffic business to Advan Research (announced November 2022) and kept its places product.
5. Map APIs. The Google Places API returns place details and search results one request at a time, so bulk analysis needs a downloadable dataset.

Placekey gives a place a two-part ID, a "what" part for the address and POI and a "where" part that encodes an H3 hexagonal grid cell at resolution 10. Overture's GERS ID is a universally unique identifier (UUID), a random code with no location inside it. It stays the same across releases and is listed in a public registry.

How POI data is collected and deduplicated

POI data collection draws on five inputs. Business websites and store locators list every location of a chain. Registries hold business filings and licenses. Social networks and review sites hold business pages. Users contribute edits, for example in OpenStreetMap and through Foursquare's Placemaker tools. Field surveys record places in person.

The hard part is conflation, also called entity resolution: deciding which records describe the same place. Overture runs each provider feed through four stages: history matching, blocking, pairwise scoring, and clustering. History matching links about 90% of a known provider's records to places already in the corpus. Blocking narrows the rest to candidate pairs, using distance and embeddings of name, category, and address. An embedding is a list of numbers in which similar text gets similar values. A gradient-boosted classifier, a model built from many small decision trees, scores each pair on name, address, website, phone, category, distance, and how distinctive the name is in its area. Matched pairs form clusters. Overture keeps one source record per cluster and merges no attributes across sources.

POI data quality and freshness

POI data changes daily as businesses open, move, and close, so quality is a property of each record. The OpenStreetMap wiki names freshness, coverage, consistency, ease of use, and customization as the marks of good POI data. Overture manages quality with these tools:

• Confidence. A score from 0 to 1 for the likelihood that a place exists. Scores are not calibrated across providers, so Overture recommends the score as a relative filter. Raising the threshold removes more junk records and also more real places.
• Operating status. Open, temporarily closed, or permanently closed, where a source reports it.
• Stable IDs and a changelog. Each monthly release publishes a data changelog that marks every ID as added, removed, changed, or unchanged. A dataset joined to GERS IDs can then be refreshed without a full re-match.

Freshness is how recently a record was checked. It is recorded in Foursquare's date_refreshed and the update_time on each Overture source record.

POI data vs foot traffic data

POI data describes a place, and foot traffic data counts visits to it. Placer.ai draws a polygon around each venue and estimates visits from aggregated, de-identified mobile device location data.

POI data examples and use cases

• Retail site selection. Retailers and restaurant chains rank candidate sites by nearby competitors, complementary businesses, and population. Gravity models, which estimate store choice from distance and attractiveness, take competitor POIs as input.
• Trade areas and cannibalization. Drive-time isochrones, the areas reachable within a set drive time, show which stores compete, as in a Wherobots cannibalization study of two Texas taco chains.
• Mobility. Visits to POIs describe how people move through a city. Epidemiologists modeled COVID-19 spread on hourly POI visits, and travel demand models use POI counts to estimate the trips each zone attracts.
• Insurance. Underwriters attach occupancy and nearby hazards to each insured location, then add up exposure for catastrophe modeling. Distance to the nearest fire station is a POI query.
• Logistics and delivery. Carriers geocode stops and measure distances with a spatial join or a nearest-neighbor search.
• Advertising and market research. Analysts count category mixes by neighborhood and link them to demographics from the census.

POI data in Wherobots

Wherobots hosts Overture places as an Apache Iceberg table (an open table format for large datasets) in the Havasu catalog: wherobots_open_data.overture_maps_foundation.places_place. A count in October 2026 returned 81,455,423 places worldwide, 17,949,232 of them with a US address. Wherobots joined Overture Maps Foundation as a contributing member in July 2024.

Overture places with isochrones holds precomputed 5, 10, 15, and 20 minute drive-time polygons for 13.3 million US places, described in the isochrones guide. Census tract boundaries, American Community Survey estimates, and Overture buildings sit in the same catalog.

Food and drink per resident in Cook County

This query counts food and drink places per 1,000 residents in each Cook County, Illinois census tract. It joins three Havasu tables in one query:

Source	Format	Size	What the query read
Overture places, places_place	Apache Iceberg table of points	81,455,423 places	32,154 food and drink places at confidence 0.7 or higher inside the Cook County bounding box
Census tracts, tiger_tract	Iceberg table of TIGER/Line polygons	All US tracts	1,332 Cook County tract boundaries
American Community Survey 5-year estimates, acs_5yr_estimates	Havasu table	All tables and geographies	1,332 tract rows of total population (B01003_001), 2020-2024

WITH tracts AS (
  SELECT GEOID, NAMELSAD, geometry
  FROM wherobots_open_data.us_census.tiger_tract
  WHERE STATEFP = '17' AND COUNTYFP = '031'
),
pop AS (
  SELECT geoid, estimate AS population
  FROM wherobots_open_data.us_census.acs_5yr_estimates
  WHERE variable = 'B01003_001' AND geo_level = 'tract' AND vintage = '2020-2024'
),
food AS (
  SELECT geometry
  FROM wherobots_open_data.overture_maps_foundation.places_place
  WHERE taxonomy.hierarchy[0] = 'food_and_drink'
    AND confidence >= 0.7
    AND bbox.xmin BETWEEN -88.27 AND -87.52
    AND bbox.ymin BETWEEN 41.46 AND 42.16
)
SELECT t.GEOID, t.NAMELSAD, p.population,
       COUNT(f.geometry) AS food_places,
       ROUND(COUNT(f.geometry) * 1000.0 / p.population, 1) AS per_1000_residents
FROM tracts t
JOIN pop p ON p.geoid = t.GEOID
LEFT JOIN food f ON ST_Contains(t.geometry, f.geometry)
WHERE p.population >= 500
GROUP BY t.GEOID, t.NAMELSAD, p.population
ORDER BY per_1000_residents DESC

The filter on bbox drops places outside the Cook County area before the point-in-polygon join, the step that tests which tract contains each point. The query ran in 18 seconds. Census Tract 8391 ranks first: 613 food and drink places for 8,355 residents, or 73.4 per 1,000. The next four tracts range from 53.1 to 49.4. With another category, or with drive-time isochrones in place of tracts, the same query screens candidate sites. The nearest-neighbor join pairs each record with its k nearest POIs.

Categories and confidence in downtown Chicago

This section profiles every Overture place inside one box: -87.66 to -87.60 longitude and 41.86 to 41.91 latitude. The box is about 5 km by 5.5 km. It covers the Loop, the West Loop, River North, Streeterville, and the South Loop. The three queries below read only the Overture places table.

WITH p AS (
  SELECT COALESCE(taxonomy.hierarchy[0], '(no category)') AS top_category, confidence
  FROM wherobots_open_data.overture_maps_foundation.places_place
  WHERE bbox.xmin >= -87.66 AND bbox.xmax <= -87.60
    AND bbox.ymin >= 41.86 AND bbox.ymax <= 41.91
)
SELECT top_category,
       COUNT(*) AS places,
       ROUND(100.0 * COUNT(*) / SUM(COUNT(*)) OVER (), 1) AS pct_of_places,
       ROUND(AVG(confidence), 2) AS avg_confidence,
       SUM(CASE WHEN confidence >= 0.7 THEN 1 ELSE 0 END) AS confidence_07_plus
FROM p
GROUP BY top_category
ORDER BY places DESC

The box holds 52,917 places, and the query ran in 6 seconds. Services and business is the largest group with 24,006 places (45.4%), followed by health care (4,451), food and drink (4,039), and shopping (3,198). 7,580 places (14.3%) have no category. They average a confidence of 0.31, against 0.79 for services and business and 0.89 for food and drink. Of the uncategorized places, 1,132 reach 0.7.

Contact fields and completeness

One pass over the same box counts the fields filled in each 0.1 confidence band:

WITH p AS (
  SELECT LEAST(FLOOR(confidence * 10), 9) AS band,
         CASE WHEN size(websites) > 0 THEN 1 ELSE 0 END AS has_website,
         CASE WHEN size(phones) > 0 THEN 1 ELSE 0 END AS has_phone,
         CASE WHEN size(websites) > 0 OR size(phones) > 0 THEN 1 ELSE 0 END AS has_website_or_phone,
         CASE WHEN size(socials) > 0 THEN 1 ELSE 0 END AS has_social,
         CASE WHEN addresses[0].freeform IS NOT NULL THEN 1 ELSE 0 END AS has_address,
         CASE WHEN brand.names.primary IS NOT NULL THEN 1 ELSE 0 END AS has_brand,
         CASE WHEN operating_status <> 'open' THEN 1 ELSE 0 END AS not_open
  FROM wherobots_open_data.overture_maps_foundation.places_place
  WHERE bbox.xmin >= -87.66 AND bbox.xmax <= -87.60
    AND bbox.ymin >= 41.86 AND bbox.ymax <= 41.91
)
SELECT CASE WHEN GROUPING(band) = 1 THEN 'all' ELSE CAST(band / 10 AS STRING) END AS confidence_from,
       COUNT(*) AS places,
       SUM(has_website) AS with_website,
       SUM(has_phone) AS with_phone,
       SUM(has_website_or_phone) AS with_website_or_phone,
       SUM(has_social) AS with_social,
       SUM(has_address) AS with_address,
       SUM(has_brand) AS with_brand,
       SUM(not_open) AS not_open
FROM p
GROUP BY ROLLUP(band)
ORDER BY confidence_from

It ran in 4 seconds. 48,177 places (91.0%) have a website or a phone number: 43,297 a phone (81.8%) and 37,998 a website (71.8%). Almost all carry a street address (52,134), 22,036 a social profile, and 3,040 a brand. 3,902 are marked temporarily or permanently closed.

Of the 8,810 places below 0.3, 8,753 have a social profile, 7,179 a phone, and 2,036 a website. Above 0.9, 84.2% have a website.

A deduplication check

Conflation leaves some duplicates behind. This self-join compares the places table with itself. It finds pairs of places within 25 m of each other with matching names. Two names match when they are equal after lowercasing and stripping punctuation. Names of eight or more characters also match when they differ by one character:

WITH p AS (
  SELECT id, geometry, confidence, sources[0].dataset AS source,
         regexp_replace(lower(names.primary), '[^\\p{L}\\p{N}]', '') AS name_key
  FROM wherobots_open_data.overture_maps_foundation.places_place
  WHERE bbox.xmin >= -87.66 AND bbox.xmax <= -87.60
    AND bbox.ymin >= 41.86 AND bbox.ymax <= 41.91
),
pairs AS (
  SELECT a.id AS id_a, b.id AS id_b,
         ST_DistanceSphere(a.geometry, b.geometry) AS dist_m,
         CASE WHEN a.name_key = b.name_key THEN 'same name' ELSE 'one character apart' END AS match_type,
         CASE WHEN a.source = b.source THEN 1 ELSE 0 END AS same_source,
         CASE WHEN a.confidence >= 0.7 AND b.confidence >= 0.7 THEN 1 ELSE 0 END AS both_07_plus
  FROM p a JOIN p b
    ON ST_DWithin(a.geometry, b.geometry, 0.0004)   -- coarse filter in degrees, about 33 m east-west
   AND a.id < b.id
  WHERE ST_DistanceSphere(a.geometry, b.geometry) <= 25
    AND length(a.name_key) >= 4
    AND (a.name_key = b.name_key
         OR (length(a.name_key) >= 8 AND levenshtein(a.name_key, b.name_key) = 1))
),
ids AS (
  SELECT id_a AS id, match_type FROM pairs UNION SELECT id_b, match_type FROM pairs
)
SELECT s.match_type, s.pairs, i.places, s.same_source_pairs, s.both_07_plus_pairs, s.median_dist_m
FROM (SELECT COALESCE(match_type, 'all') AS match_type, COUNT(*) AS pairs,
             SUM(same_source) AS same_source_pairs, SUM(both_07_plus) AS both_07_plus_pairs,
             ROUND(percentile_approx(dist_m, 0.5), 1) AS median_dist_m
      FROM pairs GROUP BY ROLLUP(match_type)) s
JOIN (SELECT COALESCE(match_type, 'all') AS match_type, COUNT(DISTINCT id) AS places
      FROM ids GROUP BY ROLLUP(match_type)) i
  ON s.match_type = i.match_type
ORDER BY s.pairs DESC

With ST_DWithin in the join condition, EXPLAIN FORMATTED shows a broadcast index join. The engine copies one side of the join to every worker and searches it through an R-tree, a spatial index of nested bounding boxes. The exact spherical distance then filters the candidates. The query ran in 66 seconds and found 58 candidate pairs covering 116 places, 0.22% of the box. 57 pairs share a name and one differs by a character. The median distance is 6.3 m. In 42 pairs both records score 0.7 or higher, and in 38 pairs the two records list different source datasets. Among the pairs are a Starbucks recorded by both AllThePlaces and Meta, and The Bagel Spot from both Foursquare and Meta. The pattern \p{L}\p{N} keeps letters and digits in any script. A pattern limited to ASCII, the basic Latin character set, strips Chinese and Korean names to empty strings and pairs unrelated records. The count is a lower bound. Names that differ by a suffix, such as "Shannon Law Group" and "Shannon Law Group, Pc", fall outside it. Production matchers such as Overture's also compare addresses, phones, and websites.

The models people use every day were trained on text, documents, databases, and the internet. Such a model can list popular restaurants in a city. Without a tool connection, it cannot count the pharmacies inside a 10-minute drive. An AI agent, a model that can call tools, can answer that with a spatial query when it is connected to the Wherobots Model Context Protocol (MCP) server.

POI data research and what changes at scale

From map icons to volunteered and platform data

Volunteered mapping and location-based social networks, both from the 2000s, produced today's open and platform POI datasets. Goodchild named volunteered geography in Citizens as sensors (GeoJournal, 2007): people with GPS devices and web maps contributing geographic facts, with OpenStreetMap as the leading case. On location-based social networks such as Foursquare, each check-in tied a person to a venue. Haklay's comparison of OpenStreetMap and Ordnance Survey data (Environment and Planning B, 2010) became an early model for judging volunteered data against an authoritative reference.

Matching and conflation

POI matching is a case of record linkage: finding the records in different datasets that describe the same thing. Fellegi and Sunter gave it a formal theory in A theory for record linkage (Journal of the American Statistical Association, 1969). Their method compares the fields of two records and weighs the agreements. It then classifies each pair as a match, a non-match, or a case for review. Places add geometry to the comparison. McKenzie, Janowicz, and Adams weighted several attribute similarities in multi-attribute matching of user-generated POIs (ACM SIGSPATIAL, 2013) and matched 97% of a sample of Yelp places to their Foursquare records. Deng, Luo, Liu, and Wang combined four attribute similarities to match Baidu and Gaode POIs in Beijing (ISPRS International Journal of Geo-Information, 2019). They used Dempster-Shafer evidence theory, a formal method for combining evidence. Low, Tekler, and Cheah built a six-step end-to-end POI conflation framework (ISPRS International Journal of Geo-Information, 2021) that reached 97.6% matching accuracy on more than 12,000 POIs.

Sun, Hu, and coauthors reviewed the field in a systematic review of POI matching methods (Computers, Environment and Urban Systems, 2023). It sorts the similarity measures and classifiers in use and lists six limitations. General entity matching moved to learned models in the same years. Mudgal and colleagues (SIGMOD 2018) found deep learning models strongest on text-heavy records and on dirty records, where values are missing or sit in the wrong field. Ditto (PVLDB, 2020) matched records with pre-trained language models. Overture applies both ideas to places with name embeddings and match labels from a large language model.

Completeness and quality

Completeness studies compare a POI source with ground truth, a reference dataset taken as correct. Klinkhardt and colleagues surveyed 49 areas in Germany and assessed OpenStreetMap POIs for travel demand models (Transportation Research Record, 2023). Completeness differed by activity type: shops were far better covered than private businesses. Yeow, Low, Tan, and Cheah catalogued 23 POI validation methods (ISPRS International Journal of Geo-Information, 2021). They applied nine of them to four datasets for one Singapore neighborhood. Spatial errors stayed under 60 m. Completeness was poor: the datasets missed real places (omission) and listed places that were not there (commission). Ballantyne and Berragan validated Overture's UK places against Geolytix supermarket locations (Environment and Planning B, 2024) and found store counts within 5% for three retailers.

POIs as features for recommendation and urban models

POI data also feeds models of how people use cities. Ye, Yin, Lee, and Lee published a POI recommendation study (SIGIR 2011). It showed that the effect of distance on check-ins follows a power law. The authors combined that distance effect with user preference and social influence. Liu and colleagues later ran an experimental evaluation of POI recommendation models (PVLDB, 2017). Yuan, Zheng, and Xie inferred the functions of city regions from POI mixes and human mobility (KDD 2012). Yan, Janowicz, Mai, and Gao learned place type embeddings with Place2Vec (SIGSPATIAL 2017) from which categories occur near each other. Huff's trading area model (Journal of Marketing, 1964) is a gravity model. Chang and colleagues modeled COVID-19 spread (Nature, 2021) on hourly movements of 98 million people between census block groups and 552,758 POIs.

What changes at scale

• Blocking becomes the bottleneck. Comparing every pair of 81 million places is out of reach, so pipelines narrow candidates by location first. In SQL that is a distance join with a spatial index. GeoSpark (Yu, Wu, and Sarwat, SIGSPATIAL 2015) and its successor Apache Sedona split that join across a cluster, as described in a GeoInformatica survey of GeoSpark (2019).
• Thresholds stop being global. A name that is rare in one town is common in another, so Overture's matcher scores name rarity within each local area.
• Identity needs upkeep. A dataset joined to GERS IDs needs the monthly changelog to stay joined, and splits and merges have no change type of their own.
• Taxonomies need crosswalks. Each source has its own categories, and Overture's mappings from provider categories are partly model-derived.

Overture's documentation names the open problems for its own data. They are duplicates, a high rate of junk records, and many empty attribute fields. The fourth is undermatching: records of one place that sit more than 100 m apart or have only moderately similar names stay unmatched.

Conclusions

The queries and studies above show that POI records vary in quality and identity by place and source, so each use of POI data needs its own checks.

1. A POI confidence score is a filter to set per task. Half of downtown Chicago places score 0.9 or higher. Below 0.3, 99.4% of records carry a social profile and 23.1% a website. The Cook County count filtered at 0.7. To keep more real places, keep the low band and mark it unverified.
2. Queries on a global POI table should filter by bounding box before the spatial join. The Cook County and Chicago profile queries ran in 4 to 18 seconds against the 81-million-row table with the bbox filter applied first.
3. Released POI data can still hold duplicates, so check every extract and read the count as a lower bound. The name-and-distance self-join found 58 pairs covering 116 places, 0.22% of the box. In 38 pairs the records list different sources. Suffix variants fall outside the check, so compare addresses, phones, and websites where duplicates change the result.
4. POI joins keyed on stable IDs can be refreshed with each release. GERS IDs carry no location and persist across releases. The changelog refreshes a join without a full re-match, except after splits and merges.
5. POI completeness is the one claim here that needs a hedge. It varies by category and place. Shops were better covered than private businesses in Germany. POI datasets for one Singapore neighborhood had errors of omission and commission. UK supermarket counts for three retailers came within 5%. Validate the category against ground truth before publishing a count.

Related terms

• Isochrone: drive-time areas around places, used for trade areas and coverage
• Spatial join: matching POIs to areas by location
• Data center site selection: screening sites by power, water, and nearby infrastructure
• Parcel data: the land records behind each address
• Building footprint: outlines of the buildings that hold places
• Ontology: classes, IDs, and relations for a POI knowledge graph
• Raster vs vector data: POIs are vector points
• Physical AI: AI that works with real places
• H3: a hexagonal grid for counting POIs per cell
• Geospatial analysis: density, proximity, and clustering methods
• Geofence: a boundary drawn around a place that triggers an event when a device enters or leaves it

Read more from Wherobots

• Location intelligence with isochrones
• Retail cannibalization study with isochrones
• How to use GERS IDs in Wherobots
• Exploring Foursquare Open Places with Wherobots

Query Overture places, census tracts, and isochrones with a Wherobots free trial at cloud.wherobots.com.