Datasets are easier to find when you provide supporting information such as their name, description, creator and distribution formats as structured data. Google's approach to dataset discovery makes use of and other metadata standards that can be added to pages that describe datasets. The purpose of this markup is to improve discovery of datasets from fields such as life sciences, social sciences, machine learning, civic and government data, and more. You can find datasets by using the Dataset Search tool.

example of Dataset Search

Here are some examples of what can qualify as a dataset:

  • A table or a CSV file with some data
  • An organized collection of tables
  • A file in a proprietary format that contains data
  • A collection of files that together constitute some meaningful dataset
  • A structured object with data in some other format that you might want to load into a special tool for processing
  • Images capturing data
  • Files relating to machine learning, such as trained parameters or neural network structure definitions
  • Anything that looks like a dataset to you

Our approach to dataset discovery

We can understand structured data in Web pages about datasets, using either Dataset markup, or equivalent structures represented in W3C's Data Catalog Vocabulary (DCAT) format. We also exploring experimental support for structured data based on W3C CSVW, and expect to evolve and adapt our approach as best practices for dataset description emerge. For more information about our approach to dataset discovery, see Making it easier to discover datasets.


Here's an example for datasets using JSON-LD syntax (preferred) in the Structured Data Testing Tool. The same vocabulary can also be used in RDFa 1.1, Microdata, or W3C DCAT vocabulary. The following example is based on a real-world dataset description.


Here's an example of a dataset in JSON-LD:


Here's an example of a dataset in RDFa:


Sites should follow the structured data guidelines. In addition to the structured data guidelines, we recommend the following sitemap and source and provenance best practices listed below.

Sitemap best practices

Use a sitemap file to help Google find your URLs. Using sitemap files and sameAs markup helps document how dataset descriptions are published throughout your site.

If you have a dataset repository, you likely have at least two types of pages: the canonical ("landing") pages for each dataset and pages that list multiple datasets (for example, search results, or some subset of datasets). We recommend that you add structured data about a dataset to the canonical pages. Use the sameAs property to link to the canonical page if you add structured data to multiple copies of the dataset, such as listings in search results pages.

Source and provenance best practices

It is common for open datasets to be republished, aggregated, and to be based on other datasets. This is an initial outline of our approach to representing situations in which a dataset is a copy of, or otherwise based upon, another dataset.

  • Use the sameAs property to indicate the most canonical URLs for the original in cases when the dataset or description is a simple republication of materials published elsewhere. The value of sameAs needs to unambiguously indicate the dataset's identity - in other words two different datasets should not use the same URL as sameAs value.
  • Use the isBasedOn property in cases where the republished dataset (including its metadata) has been changed significantly.
  • When a dataset derives from or aggregates several originals, use the isBasedOn property.
  • Use the identifier property to attach any relevant Digital Object identifiers (DOIs) or Compact Identifiers. If the dataset has more than one identifier, repeat the identifier property. If using JSON-LD, this is represented using JSON list syntax.

We hope to improve our recommendations based on feedback, in particular around the description of provenance, versioning, and the dates associated with time series publication. Please join in community discussions.

Textual property recommendations

We recommend limiting all textual properties to 5000 characters or less. Google Dataset Search only uses the first 5000 characters of any textual property. Names and titles are typically a few words or a short sentence.

Known Errors and Warnings

You may experience errors or warnings in Google's Structured Data Testing Tool and other validation systems. Specifically, validation systems may suggest that organizations should have contact information including a contactType; useful values include customer service, emergency, journalist, newsroom, and public engagement. You can also ignore errors for csvw:Table being an unexpected value for the mainEntity property.

Structured data type definitions

You must include the required properties for your content to be eligible for display as a rich result. You can also include the recommended properties to add more information about your content, which could provide a better user experience.

You can use the Structured Data Testing Tool to validate your markup.

The focus is on describing information about a dataset (its metadata) and representing its contents. For example, dataset metadata states what the dataset is about, which variables it measures, who created it, and so on. It does not, for example, contain specific values for the variables.


The full definition of Dataset is available at

You can describe additional information about the publication of the dataset, such as the license, when it was published, its DOI, or a sameAs pointing to a canonical version of the dataset in a different repository. Add identifier, license, and sameAs for datasets that provide provenance and license information.

Required properties
description Text

A short summary describing a dataset.


  • The summary must be between 50 and 5000 characters long.
  • The summary may include Markdown syntax. Embedded images need to use absolute path URLs (instead of relative paths).
  • When using the JSON-LD format, denote new lines with \n (two characters: backslash and lower case letter "n").
name Text

A descriptive name of a dataset. For example, "Snow depth in Northern Hemisphere".

Recommended properties
alternateName Text

Alternative names that have been used to refer to this dataset, such as aliases or abbreviations. Example (in JSON-LD format):

"name": "The Quick, Draw! Dataset"
"alternateName": ["Quick Draw Dataset", "quickdraw-dataset"]
creator Person or Organization

The creator or author of this dataset. To uniquely identify individuals, use ORCID ID as the value of the sameAs property of the Person type. To uniquely identify institutions and organizations, use ROR ID. Example (in JSON-LD format):

"creator": [
        "@type": "Person",
        "sameAs": "",
        "givenName": "Jane",
        "familyName": "Foo",
        "name": "Jane Foo"
        "@type": "Person",
        "sameAs": "",
        "givenName": "Jo",
        "familyName": "Bar",
        "name": "Jo Bar"
        "@type": "Organization",
        "sameAs": "",
        "name": "Fictitious Research Consortium"
citation Text or CreativeWork

Identifies academic articles that are recommended by the data provider be cited in addition to the dataset itself. Provide the citation for the dataset itself with other properties, such as name, identifier, creator, and publisher properties. For example, this property can uniquely identify a related academic publication such as a data descriptor, data paper, or an article for which this dataset is supplementary material for. Examples (in JSON-LD format):

"citation": ""
"citation": ""
"citation": ""
 "Doe J (2014) Influence of X ..."

Additional guidelines

  • Don’t use this property to provide citation information for the dataset itself. It is intended to identify related academic articles, not the dataset itself. To provide information necessary to cite the dataset itself use name, identifier, creator, and publisher properties instead.
  • When populating the citation property with a citation snippet, provide the article identifier (such as a DOI) whenever possible.

    Recommended: "Doe J (2014) Influence of X. Biomics 1(1)."

    Not recommended: "Doe J (2014) Influence of X. Biomics 1(1)."

identifier URL, Text, or PropertyValue

An identifier, such as a DOI or a Compact Identifier. If the dataset has more than one identifier, repeat the identifier property. If using JSON-LD, this is represented using JSON list syntax.

keywords Text

Keywords summarizing the dataset.

license URL, CreativeWork

A license under which the dataset is distributed. For example:

"license" : ""
"license" : {
  "@type": "CreativeWork",
  "name": "Custom license",
  "url": ""

Additional guidelines

  • Provide a URL that unambiguously identifies a specific version of the license used.


    "license" : ""

    Not recommended

    "license" : ""
sameAs URL

URL of a reference Web page that unambiguously indicates the dataset's identity, usually in a different repository.

spatialCoverage Text, Place

You can provide a single point that describes the spatial aspect of the dataset. Only include this property if the dataset has a spatial dimension. For example, a single point where all the measurements were collected, or the coordinates of a bounding box for an area.


"spatialCoverage:" {
  "@type": "Place",
  "geo": {
    "@type": "GeoCoordinates",
    "latitude": 39.3280,
    "longitude": 120.1633


Use GeoShape to describe areas of different shapes. For example, to specify a bounding box.

"spatialCoverage:" {
  "@type": "Place",
  "geo": {
    "@type": "GeoShape",
    "box": "39.3280 120.1633 40.445 123.7878"

Points inside box, circle, line, or polygon properties must be expressed as a space separated pair of two values corresponding to latitude and longitude (in that order).

Named locations

"spatialCoverage:" "Tahoe City, CA"
temporalCoverage Text

The data in the dataset covers a specific time interval. Only include this property if the dataset has a temporal dimension. uses the ISO 8601 standard to describe time intervals and time points. You can describe dates differently depending upon the dataset interval. Indicate open-ended intervals with two decimal points (..).

Single date

"temporalCoverage" : "2008"

Time period

"temporalCoverage" : "1950-01-01/2013-12-18"

Open-ended time period

"temporalCoverage" : "2013-12-19/.."
variableMeasured Text, PropertyValue

The variable that this dataset measures. For example, temperature or pressure.

version Text, Number

The version number for the dataset.

url URL

Location of a page describing the dataset.


The full definition of DataCatalog is available at

Datasets are often published in repositories that contain many other datasets. The same dataset can be included in more than one such repository. You can refer to a data catalog that this dataset belongs to by referencing it directly.

Recommended properties
includedInDataCatalog DataCatalog

The catalog to which the dataset belongs.


The full definition of DataDownload is available at In addition to Dataset properties, add the following properties for datasets that provide download options.

The distribution property describes how to get the dataset itself because the URL often points to the landing page describing the dataset. The distribution property describes where to get the data and in what format. This property can have several values: for instance, a CSV version has one URL and an Excel version is available at another.

Required properties
distribution.contentUrl URL

The link for the download.

distribution DataDownload

The description of the location for download of the dataset and the file format for download.

distribution.encodingFormat Text, URL

The file format of the distribution.

Tabular datasets

A tabular dataset is one organized primarily in terms of a grid of rows and columns. For pages that embed tabular datasets, you can also create more explicit markup, building on the basic approach described above. At this time we understand a variation of CSVW ("CSV on the Web", see W3C), provided in parallel to user-oriented tabular content on the HTML page.

Here is an example showing a small table encoded in CSVW JSON-LD format. There are some known errors in the Structured Data Testing Tool.

Help and tools