What the web actually marks up

Key findings

• 46.68% of the 2,487 real Schema.org terms are used by fewer than 1,000 websites.
• The figure most sites republish, 76.05%, counts 3,058 annotation rows that are not vocabulary.
• Only 51 terms are on ten million websites or more.
• Google publishes six usage bands and no exact count inside any of them.
• The vocabulary Google measures holds 937 classes, 1,529 properties, and 20 enumeration values.
• Three sources measure structured data on the web, and no two of their figures can be combined.

46.68% of terms sit under 1,000 websites, not 76.05%

July 2026 · six ranges, no exact numbersGoogle and the Schema.org community, Schema.org usage statistics dataset. Websites in Google's index.

The two bars show the same measurement counted two ways. The dashed bar is the wrong way.

The file holds 5,545 rows, but 3,058 of them end in -input or -output. Those are markers from Schema.org's Actions syntax, two per real property, and you would not write one into a page. Almost all of them sit in the smallest band, so leaving them in drags the whole distribution downwards and makes genuinely rare terms look ordinary. I expected the correction to move the headline by a point or two. It moved it by more than that. If the figure you have seen elsewhere is the higher one, the markers are the whole difference.

Only 51 terms are on ten million websites or more

July 2026 · six ranges, no exact numbersGoogle and the Schema.org community, Schema.org usage statistics dataset. Websites in Google's index.

The bars count terms in Google's six usage bands for July 2026.

Nearly half of the vocabulary sits in the smallest band. The counts fall away fast above it. Adoption concentrates in a small set of terms the whole web shares. That is the shape I expected. The steepness still surprised me. For your own markup it means the safe terms are few and well known. Anything outside them puts you in thin company.

The labels are ranges, because Google does not publish the exact figure inside a band.

Three measurements, three different webs

The dateline marks when the three sources measured. It shows no quantities at all, on purpose.

Three sources, three webs. Google measures its whole index and publishes ranges. The Web Almanac measures about 15 million origins once a year and publishes exact counts. Web Data Commons measured a Common Crawl corpus and stopped in December 2024. The one I lean on is Google, because it is the only one that covers the whole vocabulary. The one I would reach for to sanity-check a single term is the Web Almanac. If you see a page blend the two into one number, nobody measured that number.

I do not combine any two of these into one figure, because an exact count of one set of websites is not the number hiding inside a range measured across another.

Reports

• The properties Google recommends and almost nobody uses103 of 206 recommended properties are on fewer than 100,000 websites.
• Every Schema.org term, and how much the web uses itAll 2,487 terms in one table you can search and sort, with Schema.org's own definitions.
• What each industry can mark up, and what it actually does578 of the 937 classes sit in a real industry, and four of the thirteen industries have no rich result Google documents at all.
• What changed, and how long adoption takes136 terms ended three archived months in a different usage band, 87 have been superseded with a documented replacement, and 22 are already on 100,000 websites before Schema.org has ratified them.
• The properties Google requires, and what they rest on30 of 32 live features need a required property, and one of them rests on a property used by fewer than 1,000 websites.
• Nearly a third of the type names on the web are not real209 of the 682 distinct type strings the Web Almanac's 2025 crawl found are not real Schema.org types, 30.65% of everything written.
• Where the vocabulary lives, and what makes a term catch onHealth is the largest extension field Schema.org publishes and the least used, 75.00% of its terms under a thousand websites. Properties attached to more types get used more, without exception.
• How these figures are madeThe five sources and their licences, the rules the pipeline enforces, and three mistakes I found in my own published figures and fixed.

Questions

How many Schema.org terms are actually used on the web?

Google's July 2026 file covers 2,487 real terms. That breaks down into 937 classes, which are the types you mark something up as, 1,529 properties, which are the fields inside them, 20 enumeration values, and one term Schema.org has not actually published. 46.68% of those terms appear on fewer than a thousand websites. Only 51 of them reach ten million or more.

Why does this page show a different percentage from other sites?

Because 3,058 rows in the file are not vocabulary terms at all. They end in -input or -output, two per real property, and they come from Schema.org's Actions syntax, not from anything you would mark up. Leave them in and you get 76.05% of terms below a thousand websites, which is the figure you will find published elsewhere. Take them out first and you get 46.68%. You can check this yourself against the file. That is why I publish both numbers and not only the corrected one.

Are these exact numbers of websites?

No, and nobody has them. Google publishes six bands, not exact counts, so a term is known to sit somewhere between one thousand and ten thousand websites without the real figure ever being made public. If you find a page showing an exact count for Google's data, it has either been invented or taken from a different measurement of a different set of websites.

Where does the data come from?

Google publishes it monthly in the Schema.org repository. I keep a checksummed copy of the files, so the record survives even if the source stops. The rich result requirements are separate. I read those from the markup of Google Search Central's own pages on 11 August 2026, not from memory, and each one carries the address and date it came from.

Can I reuse these numbers?

Yes. You can download the full dataset as JSON or CSV at a stable address, and the analysis is licensed under the Apache License 2.0. All I ask is a credit to eduarddziak.com with a link. One source works differently. Web Data Commons publishes no licence for its data at all, so I cite and link its figures and keep them out of the downloads.

How often does this update?

Google publishes a new file monthly, and I archive it, check it against the previous months, and rebuild the figures from the archive. You are looking at July 2026, rebuilt on 19 August 2026.