01 / Overview

Discover Areas With Similar Profiles

Population Dynamics Insights represents geographic areas as numerical profiles built from aggregated signals. Compare areas to find similarities, or evaluate these profiles as inputs to your own models.

See What's Inside Vs. Places Insights

From Geography To Similarity

Conceptual Overview

Geography

An S2 Cell

A geographic unit in the supported dataset.

Representation

330 Dimensions

A learned embedding for the cell.

Comparison

Similar Areas

Compare embeddings to identify areas to investigate.

Similarity describes resemblance. It does not establish demand, revenue, or expected site performance; predictive use requires independent outcome validation.

At A Glance

Provides: Geographic embeddings for similarity analysis and model development.
Not: Raw searches, individual movement records, or validated business predictions.

330dimensions per embedding

12S2 level for the shown dataset

3documented source blocks

17country datasets in the setup list

About PDI data

02 / What's Inside

Three Source Blocks, One Learned Representation

The documented vector has three blocks: aggregated Search Trends, Maps and busyness, and weather/air quality. Individual dimensions are latent features rather than named measurements. Comparing models with and without a block can help test its contribution to a defined task.

0–127 128–255 256–329

Aggregated Search Trends — learned regional patterns from aggregated search signals; not individual intent.

Maps and Busyness — learned features from place context and aggregate activity density.

Weather and Air Quality — environmental context encoded in latent features.

Feature ablation can compare a block’s contribution under your evaluation design. It does not by itself establish why a location performs well.

About PDI data

Built For Power

one embedding per covered geographic cell for similarity analysis
known source blocks; individual dimensions remain abstract

Kept Private

no raw search queries or device-level mobility traces
no individual-level data of any kind

03 / Data & Schema

Join Your Locations To Geographic Cells

Subscribe to the relevant country dataset in BigQuery. Convert your location coordinates into S2 level-12 cell tokens, then join them to geo_id. Validate coverage and the geographic/time alignment of your own data before comparing or modeling locations.

Key Facts

geo_id is an S2 cell token; convert point coordinates to the matching level and string representation
Administrative-area identifiers are Google Place IDs; do not confuse a cell with an individual POI
Use the subscribed dataset’s actual coverage and version; a missing cell is not evidence of no activity

Core schema reference PDI overview docs

Real-world signalsaggregated source signals

Foundation modellearns one embedding per S2 cell

Population Dynamics Insights (BigQuery)country-specific S2 level-12 dataset

Joins your BigQuery dataon a matching S2 token; prepare and validate your own data

BigQuery MLVertex AIYour GCP data

population_dynamics___us.v1_s2_z12Country dataset

geo_idSTRING

featuresFLOAT64[330]

administrative_area_level_1_idSTRING

administrative_area_level_2_idSTRING

···· one embedding, no named columns

04 / Example Query

Compare Similarity, Then Investigate

These worked examples show cosine similarity across full embeddings and across the Search Trends block. Replace the project and target token with your own linked dataset and a covered cell. Output is synthetic, not measured query output; a similarity score is not a performance forecast.

Full Embedding

Synthetic identifiers and scores; exclude the target cell

WITH target AS (
 SELECT geo_id, features
 FROM `YOUR_PROJECT.population_dynamics___us.v1_s2_z12`
 WHERE geo_id = '80ead45'
)
SELECT t.geo_id,
 ROUND(1 - ML.DISTANCE(t.features, x.features, 'COSINE'), 3) AS similarity
FROM `YOUR_PROJECT.population_dynamics___us.v1_s2_z12` AS t
CROSS JOIN target AS x
WHERE t.geo_id != x.geo_id
ORDER BY similarity DESC
LIMIT 4

Illustrative output · No request executed

geo_id	Similarity
example_cell_a	0.921
example_cell_b	0.904
example_cell_c	0.896
example_cell_d	0.881

Search-Block Comparison

Rankings can change when the feature block changes

WITH sliced AS (
 SELECT geo_id, ARRAY_SLICE(features, 0, 127) AS search_block
 FROM `YOUR_PROJECT.population_dynamics___us.v1_s2_z12`
), target AS (
 SELECT geo_id, search_block FROM sliced
 WHERE geo_id = '80ead45'
)
SELECT s.geo_id,
 ROUND(1 - ML.DISTANCE(s.search_block, t.search_block, 'COSINE'), 3) AS similarity
FROM sliced AS s CROSS JOIN target AS t
WHERE s.geo_id != t.geo_id
ORDER BY similarity DESC
LIMIT 4

Illustrative output · No request executed

geo_id	Similarity
example_cell_e	0.933
example_cell_b	0.919
example_cell_g	0.907
example_cell_h	0.894

Use the PDI embeddings

05 / Analysis Modes

Choose An Analysis And Its Validation

Use similarity and clustering to explore geographic patterns. For supervised prediction, bring outcome labels, define the task, and evaluate on held-out locations and periods. Sparse observations do not automatically become reliable estimates.

Cluster labels are analyst interpretations. Validate them against independent context; the vector does not provide demographic or lifestyle labels.

Use the PDI embeddings

Find Comparable AreasRank covered cells by vector similarity.

Use the result as a shortlist for investigation; compare market context before acting.

Test a Supervised ModelAdd embeddings to a model with observed outcomes.

Compare with a baseline and hold out locations. Report error and uncertainty; a few examples may be insufficient.

Evaluate Geographic TransferTest performance in a new region.

Hold out entire markets to detect geographic leakage and distribution shift before using the model there.

Explore Spatial DisaggregationTest finer-grained estimates against independent data.

A model can allocate a coarse statistic; it does not create observed neighborhood measurements.

Evaluate Over TimeUse a temporal evaluation with appropriate data.

An embedding is not a forecast. Validate feature dates and outcomes with time-based holdouts.

06 / Use Cases

Use Embeddings As Evidence To Test

Explore comparable areas, evaluate candidate locations with additional evidence, and test whether embeddings improve an existing model. Each use needs a defined geography, time period, outcome, and validation method.

Compare AreasFind similar geographic profiles, then check the context independently.

Evaluate a CandidateCombine similarity with site facts and observed outcomes; do not infer expected revenue from resemblance.

Describe a ClusterUse documented place context to interpret a cluster; do not infer personal traits.

Test Model ImprovementCompare accuracy with and without embeddings on held-out data; improvement is not guaranteed.

Potential applications include retail, public health, and planning. Validate each application against its own outcomes and constraints.

Industries

Retail & QSR Real estate Financial services Insurance Public health Government & planning Franchising Hospitality & travel Logistics & delivery Telecom CPG & consumer goods Automotive Media & entertainment

07 / Privacy And Explainability

Separate Representation From Explanation

The dataset provides aggregate-derived latent features, not raw search histories or mobility traces. Explain a similarity result with independently sourced geographic context. An LLM can summarize supplied evidence but cannot reveal the meaning of individual dimensions or turn resemblance into a causal explanation.

Privacy-safe By Design

No raw individual search histories or device-level movement records
The embedding represents an area, not a person

Bring It To Life

Use Places Insights or other documented context to describe an area
Keep interpretation, prediction, and validation distinct

About PDI data

Evidence And Interpretation

Representation

330 latent features for a covered geographic cell

Comparison

A similarity score for a defined vector comparison

Context

Independent, dated geographic facts that help interpret the result

Validation

Observed outcomes and held-out testing for any predictive claim

Similarity alone does not establish demand, causality, or future performance.

08 / Vs. Places Insights

Two Different Questions, Two Different Answers

Places Insights supplies structured place attributes through governed analysis. PDI supplies learned geographic features for comparison and modeling. They answer different questions and can complement one another when joined at the correct geography.

Dimension	PDI	Places Insights
Representation	Learned geographic patterns	Structured attributes of places
Shape	330-dimensional vector per cell	Named fields; governed query outputs
Geography	S2 cell; area varies geographically	Place-level source, constrained outputs
Interpretation	Latent features; not named measures	Fields such as type, rating, and hours
Typical Use	Similarity exploration and ML features	Place counts, mix, and aggregate analysis
Access Rules	Check the country listing and terms	SQL threshold; separate count-function rules
Prediction	Requires a separately validated model	Requires outcomes and a validated model

The pairing is the point: PDI finds where to look, Places Insights explains what's there once you've found it.

09 / Coverage And Access

Select The Country Dataset And Verify Coverage

The current setup documentation lists 17 country datasets. Select the relevant listing, complete onboarding, and check available cells and dataset version. Access to one country does not establish coverage or entitlement elsewhere.

Coverage

Country-specific datasets share the documented embedding schema
Check cell availability and geographic alignment for your use case

Access

Follow Google’s current signup and cloud-setup instructions
Budget for applicable dataset access and BigQuery usage

Recheck country availability and launch-stage terms before quoting or committing to an implementation.

Country Setup and Access

Start With The Dataset

Select

Choose an available country listing.

Prepare

Link the dataset and map your coordinates to cell tokens.

Evaluate

Check coverage, compare a baseline, and validate the intended use.

Conceptual diagram