Dataset Markup Spread to 5.5 Times as Many Top-1M Sites Since 2023, Yet Just 13 Government and University Sites Use It

Summary

Picture a county health department that publishes a spreadsheet of restaurant inspection scores. A local reporter finds it, but an AI assistant asked "which restaurants failed inspection this year" may never connect the question to that file, because nothing on the page says it is a dataset, what it covers or where to download it.

Dataset markup does exactly that. It names the dataset, describes it, states its license, the period and place it covers, and links to the files. Google uses it for Dataset Search, and it gives AI engines a clean description of data they might cite.

Adoption among top sites has grown quickly from a small base, from 84 sites in 2023 to 459 in 2026. Government and university sites, which publish much of the world's open data, have barely started.

What we measured

We used four October crawls from HTTP Archive (2023 to 2026), mobile client, top-1M sites. HTTP Archive crawls each site's homepage plus one inner page. We parsed the rendered JSON-LD on each page, flattened nested objects and @graph arrays, and flagged pages with a Dataset node.

We grouped sites by host. Government means hosts ending in .gov, .mil, a country government suffix such as .gov.uk, .gouv.fr or .gob.mx, .gc.ca or .europa.eu. Education means .edu, .edu.xx or .ac.xx. For 2026 we counted which properties the Dataset nodes carry.

Dataset markup grew fast from a small base

Growth was steady for two years and then more than doubled in the last twelve months.

Government and education sites made up 2 of 84 adopters in 2023 and 13 of 459 in 2026.

Government and universities have barely started

These sites publish large amounts of data, but their crawled pages rarely describe it in markup.

Only one inner page per site is crawled, so a portal with thousands of dataset pages can still be missed. Sampling affects every group the same way, though, and the combined rate for government and education sites is under half that of other sites.

Most Dataset markup skips the details AI engines need

Names and descriptions are almost universal. Licensing and coverage are not.

A dataset with no license, coverage or download link is hard for any machine to recommend with confidence.

What this means

If your business publishes data, such as price indexes, survey results, market reports or local statistics, give each dataset its own page with Dataset markup. Include name, description, license, creator, temporalCoverage, spatialCoverage, variableMeasured and a distribution with the file URL and format. It takes about an hour per dataset, and it is the sort of detail AI SEO services should handle for you.

Original data is one of the things AI engines cite most readily, and well-described data is easier to cite. If assistants do not yet reference your reports, better markup is one of the simpler AI search optimization fixes to make, and a free AI visibility audit will show you where they stand.

Limits

HTTP Archive crawls one homepage and one inner page per site. Dataset pages usually sit deep in a site, so these counts show adoption in the sample, not the true number of sites with dataset markup.

Government and education detection uses domain suffixes and misses agencies or universities on .org or .com domains.

The 2026 property counts for government and education rest on 16 pages, too few to break out separately.

The top-1M site list changes between crawls.

Frequently asked questions

Related studies

• Niche schema types AI answers can read1 in 68 JSON-LD Blocks on Top Sites Fails to Parse
• Niche schema types AI answers can readService Markup Spread to 5x More Top Sites Since 2023
• Niche schema types AI answers can read87.6% of Forum Pages Lack Discussion Markup
• Niche schema types AI answers can read98% of Course Creator Sites Show No Course Markup