Dataset schema: getting research data listed in Google Dataset Search

JSON-LD example

Copy this into a <script type="application/ld+json"> tag and replace the values with your own.

{
  "@context": "https://schema.org",
  "@type": "Dataset",
  "name": "Riverside Allotment Soil Tests, 2021-2025",
  "description": "Soil pH, organic matter and lead readings from 64 community garden plots, sampled each spring and autumn by volunteers and analyzed by a university extension lab.",
  "url": "https://example.com/data/allotment-soil-tests/",
  "identifier": "https://doi.org/10.1234/example.soil.2025",
  "keywords": [
    "soil chemistry",
    "urban agriculture",
    "lead contamination"
  ],
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "isAccessibleForFree": true,
  "version": "3",
  "creator": {
    "@type": "Organization",
    "name": "Riverside Growers Association",
    "url": "https://example.com/"
  },
  "temporalCoverage": "2021-04-01/2025-10-31",
  "spatialCoverage": {
    "@type": "Place",
    "name": "Riverside Allotments, Dayton, Ohio"
  },
  "distribution": [
    {
      "@type": "DataDownload",
      "encodingFormat": "CSV",
      "contentUrl": "https://example.com/wp-content/uploads/2025/11/soil-tests-2021-2025.csv"
    }
  ]
}

Properties

What counts as a dataset

Google's definition is wide. A table or a CSV file qualifies, and so does an organized collection of tables, a proprietary file that holds data, a group of files that only make sense together, images that capture data, and machine learning files such as trained parameters.

The markup describes the data, not the data itself. It says what the dataset is about, who made it, which period and place it covers and where to download it. It never lists the actual values.

Put the markup on the dataset's landing page: the one page that exists to describe it. Repositories also have listing and search pages that mention many datasets. Those don't need markup, and if you add it there anyway, point sameAs at the landing page so Google knows which description is canonical.

Republished and derived data

Open data gets copied, cleaned and combined constantly, and Google asks you to say which case you are in:

• A straight republication of someone else's dataset: add sameAs with the original's canonical URL. Each sameAs value must identify one dataset only.
• A version you changed significantly, or one built from several sources: use isBasedOn for each original.
• Persistent identifiers: attach DOIs or Compact Identifiers with identifier, repeated as a list if there are several.

Keep text properties under 5000 characters, because Dataset Search ignores anything past that point. Names and titles should be a few words, not a paragraph.

Publishing data from a WordPress site

A common setup is a page that explains the dataset plus the files uploaded to the Media Library. The page is the landing page; each attachment URL becomes a contentUrl in distribution.

Hydrogen SEO's schema screens don't offer a Dataset type, so print it with the hydrogen_seo_schemas filter, which receives every schema object before output:

add_filter( 'hydrogen_seo_schemas', function ( $schemas ) {
    if ( is_page( 'allotment-soil-tests' ) ) {
        $schemas[] = [
            '@context'    => 'https://schema.org',
            '@type'       => 'Dataset',
            'name'        => 'Riverside Allotment Soil Tests, 2021-2025',
            'description' => get_the_excerpt(),
            'license'     => 'https://creativecommons.org/licenses/by/4.0/',
            'distribution' => [ [
                '@type'          => 'DataDownload',
                'encodingFormat' => 'CSV',
                'contentUrl'     => wp_get_attachment_url( 812 ),
            ] ],
        ];
    }
    return $schemas;
} );

Swap in your page slug and attachment ID. To pull a dataset out of Dataset Search later, add a robots noindex to the landing page; Google says the change can take days or weeks to show.

Warnings you can ignore, and ones you can't

Google lists two known quirks. Validators may insist that an Organization in creator needs contact details with a contactType (useful values include customer service, emergency, journalist, newsroom and public engagement), and the Rich Results Test may complain that csvw:Table is an unexpected value for mainEntity. The second one is safe to ignore. The CSVW approach for embedding tabular data is itself still in beta.

Real errors are different: a missing description, one under 50 characters, or a contentUrl that returns 404. Check syntax with the JSON-LD checker, then open every download link by hand.