The Shape of Schema.org at Scale

01. Corpus

Seven contiguous crawls, one analytical shape.

The corpus consists of every monthly Common Crawl snapshot released in the first seven months of 2026: CC-MAIN-2026-04 (January), 08 (February), 12 (March), 17 (April), 21 (May), 25 (June) and 30 (July). Each crawl contains 100,000 WARC files. Across the seven months, 700,000 WARC files were processed successfully. 127 files in the March crawl returned S3 throttling errors and were treated as zero-page entries in per-file statistics but excluded from coverage histograms.

02. Coverage

The web got more machine-readable every month.

Weighted mean coverage across the full corpus is 58.32%. Per-month coverage rises monotonically from a low of 55.27% in February to a high of 60.05% in July, an increase of 4.78 percentage points over the seven months. January registers 57.31%, then February dips to 55.27% before the monthly figure rises through March (57.69%), April (59.00%), May (59.19%), June (59.73%) and July (60.05%). The February dip is a real crawl-composition effect rather than a data-quality artefact: the aggregation was re-run against the raw per-file logs to confirm no error mass was hidden.

Per-file variance is remarkably tight. The standard deviation of coverage across the 100,000 WARC files of each month is below 1.25 percentage points in every month, and below 0.75 percentage points in five of the seven. Distinct-type counts per WARC file cluster around a weighted mean of 509 with a standard deviation of approximately 21 in each month. This uniformity implies that the underlying phenomenon, which pages carry schema markup and how many types they carry, does not depend strongly on which subset of the crawl is sampled.

03. Distribution shape

A hundred types carry almost all of it.

Vocabulary usage is extremely concentrated. Table 2 gives the cumulative percentage of total occurrences accounted for by the top N types, using the June 2026 crawl as a representative month; the shape is essentially unchanged in every other month. A single type, ListItem, accounts for approximately one in every six occurrences in every month examined, between 16.67% and 17.05%.

The long tail is correspondingly heavy. In each month, more than 62% of distinct types are singletons, appearing exactly once. In January the singleton share peaks at 64.9%, 114,765 out of 176,912 distinct types. Types appearing 10 times or fewer account for approximately 85% to 90% of distinct types in every month, and types appearing 100 times or fewer for approximately 95%.

04. Category composition

The web marks up its furniture.

Grouping the top 200 types per month into ten fixed categories yields a picture that is stable across all seven months: no category shifts by more than 3.1 percentage points across the window, and the overall shape is dominated by structural markup. Navigation and Structure sits between 43.5% and 46.6% of top-200 occurrences across the seven months, with a slight downward trend from 46.6% in January to 43.5% in July. It is populated overwhelmingly by ListItem, BreadcrumbList, SiteNavigationElement, WebPage and WebSite.

Content and Media sits stably at approximately 18%, driven by ImageObject at 1.6 to 2.0 billion occurrences per month, Article and blogging types. E-commerce (Offer, Product, Brand, Rating, Review) rises slightly across the seven-month period, from 11.2% to 12.6%, consistent with growing commercial schema adoption. Places and Geo rises from 5.2% to 6.2% over the same period. Events and Actions, Interaction and Social, Vehicles, and Technology and Software collectively contribute less than 1.5% of top-200 occurrences in every month. This is not because the underlying types are rare in absolute terms, since Vehicles alone accounts for tens of millions of occurrences per month, but because they are dwarfed by the structural and content categories.

05. Format evolution

Microdata is losing, one month at a time.

The stratified 200-WARC per-month sample shows a consistent, monotonic migration from microdata toward JSON-LD across the seven-month period. JSON-LD page share rises from 74.53% in January to 77.84% in July, a gain of 3.31 percentage points. Microdata page share falls from 49.40% to 45.14% over the same period, a loss of 4.26 percentage points. Both directions are monotonic: every month shows JSON-LD ticking up and microdata ticking down relative to the prior month.

The overlap between the two, that is, the percentage of pages carrying both formats, is stable throughout at approximately 23% to 24%: 22.98% in July, 23.95% in May, 23.93% in January. The migration is not therefore a matter of dual-format pages dropping one; it is microdata-only pages converting to JSON-LD or being replaced with JSON-LD-only equivalents. The shift is more pronounced when measured by occurrences rather than pages: JSON-LD's share of total type occurrences rises from 70.49% in January to 75.66% in July, meaning JSON-LD-carrying pages are becoming denser in type occurrences relative to microdata pages as well as more numerous.

06. Shadow layer

Eight hundred million occurrences a parser would reject.

Combining the four error categories yields a stable shadow layer representing invalid or non-conforming markup. It represents 0.63% of all type occurrences across the seven-month corpus: 811,946,273 out of 127,990,580,002. The per-month share ranges from 0.58% in July to 0.67% in January, a spread of 0.09 percentage points. No month exceeds the 0.5-percentage-point variance threshold from the mean, and there is a weak but consistent downward trend consistent with improving implementation quality over time.

Property-as-type errors are the largest single category at 49.2% of the shadow layer. Because JSON property casing conventionally begins with a lowercase letter and schema.org type casing conventionally begins with uppercase, this error is easily detectable and has a clear rewriting fix. Two error patterns emerged as high-frequency in July that had not been present at that scale in earlier months: ImageObject with a zero substituted for the letter O, at approximately 5.6 million occurrences, and http//schema.org/Review missing the colon in the URL scheme, at approximately 202 thousand occurrences. Both appear to have propagated through a single templating library or CMS plugin update.

07. Co-occurrence templates

Seven bundles the web ships over and over.

Co-occurrence is measured on a stratified sample of 10 WARC files per month, on the schema-bearing pages within them, approximately 115,000 pages per month. For each ordered pair of types (A, B) the analysis computes P(B | A) and P(A | B) by counting the pages on which each type appears. Highly correlated pairs are those where both conditionals exceed 80%; asymmetric pairs are those where P(B | A) is at least 90% but P(A | B) is below 40%. Every month yields the same seven templates, at conditional probabilities above 87%.

Figure 6. Seven co-occurrence templates. All seven hold across all seven months at the thresholds shown; no template fell below the 80%/80% symmetric threshold in any month for its core pair. Node colour marks the functional category, and neutral nodes mark rider types.

08. Wikidata linking

One page in 384 grounds itself in a knowledge base.

A separate full-corpus scan of the June 2026 crawl (CC-MAIN-2026-25) only was run to measure how often schema.org markup carries a link to a Wikidata entity via the sameAs, mainEntity, mainEntityOfPage or @id property. The run examined 2,065,085,607 pages across 100,000 WARC files. Only 0.261% of all pages, about one in every 384, carry a Wikidata reference. Among schema-bearing pages the rate is higher but still small.

This finding matters because Wikidata linking is the most direct signal that a page is not just marking up its own content, but is grounding the entities on that page in a shared, resolvable knowledge base. QID concentration is heavy but less extreme than schema.org type concentration: the top 1 QID accounts for 1.06% of QID occurrences, the top 100 for 18.86%, the top 1,000 for 40.75% and the top 10,000 for 62.48%.

09. Comparison

Two orthogonal signals that broadly agree.

Google's Schema.org Usage Statistics Dataset, released June 2026, reports the number of distinct root domains observed to use each schema.org type. This study reports the number of instance occurrences. These are two orthogonal signals: a type used ten million times on a handful of large sites will rank high in this study and low in Google's, while a type used a handful of times on a million sites will rank the opposite way.

The two signals nonetheless broadly agree on which schema.org types are most prominent. Spearman rank correlation on the top 100 types is 0.86, using instance ranking as one signal and domain ranking as the other. Types such as Organization, WebSite, WebPage, BreadcrumbList, ImageObject, Person and ListItem appear in the top tier of both rankings. The disagreements are informative: types associated with a small number of very high-traffic sites, for example template-heavy SiteNavigationElement usage, rank higher on instance volume than on domain breadth, while types associated with wide but shallow adoption rank higher on domain breadth than on instance volume.

The domain-level analysis, a stratified sample of 500 WARC files per month, shows a 45.8% increase in distinct schema-bearing domains from January (132,525) to July (193,179), and confirms that the top seven types by domain breadth (ListItem, Organization, BreadcrumbList, ImageObject, Person, WebPage, WebSite) are the same in every month and match Google's highest domain-breadth tier exactly. From March 2026 onward, Offer and Product enter the top 10 by domain count, an e-commerce adoption signal that is visible in domain breadth before it fully registers in instance volume.

10. Method

Methodology.

11. Limitations

Two limitations should be flagged.

12. References

References.