Entity SEO: 107 of 163 schema topic statements named the site itself

What is entity SEO, and what does it ask your markup to say?

It asks for two separate statements and most sites only make the first. The first is identity: this site belongs to this organisation, which has this name, this logo and these profiles elsewhere. That one is well covered, and we have measured it from several directions: 412 of 1,080 home pages named their site in neither source Google reads first, and 85 of 317 organizations named a reference entry in their sameAs list.

The second statement is subject matter, and it is a different claim. An engine that has resolved who you are still does not know what any given page of yours is for. Four properties in the vocabulary carry that information, and they divide the work cleanly. About names the subject of a document. MainEntity, defined as the primary entity described in some page or other CreativeWork, names the one thing a page is principally a page about, and its inverse is mainEntityOfPage. KnowsAbout is the odd one out: it hangs off an Organization or a Person rather than a document, and the specification is careful to say it indicates a topic that is known about, suggesting possible expertise but not implying it. IsPartOf, defined as an item that this item is part of, places a document inside a larger one rather than describing it.

The distinction matters because only the first two are statements about the page in front of the crawler. The counts below keep them apart for that reason, and they also exclude one case that would otherwise inflate mainEntity badly. On a FAQPage or a QAPage, mainEntity does not hold a topic: it holds the questions. We found 51 such nodes on 50 sites and took them out of every mainEntity figure in this post, because counting them would mean counting the FAQ markup we have measured separately as a topic statement.

How many sites named a subject at all?

129 of the 612, which is one in five of the pages that had already gone to the trouble of publishing structured data. The denominator is worth holding on to. These are not sites that ignore markup. They had 1,151 parseable JSON-LD blocks between them holding 8,731 nodes, which is a mean of just over fourteen nodes a page. They are the engaged minority. We have measured the other side of that line before: 141 of 382 home pages carried no structured data at all in the raw HTML.

Within the engaged minority the four properties are not adopted evenly, and the ordering is the opposite of what their definitions would suggest. The property declared most often is isPartOf, on 138 sites, which says the least: it places a page inside a site without describing either. The property declared least often is mainEntity, on 18 sites, which says the most. About sits between them on 97 sites. A vocabulary where the most informative property is the least used is a vocabulary being filled in by something other than an author with an intention.

The strata make the same point from another angle. Among the 89 software-as-a-service hosts carrying JSON-LD, 23 named a subject. Among the 52 news hosts, 13 did. Among the 26 government hosts, one did. The pages least likely to say what they are about were the ones published by institutions whose whole output is subject matter, and the pages most likely to say it were commercial sites whose subject is usually themselves. That is a hint about where these properties come from, and the next section turns it into a count.

107 of 163 topic statements named the publisher, not a topic

This is the measurement the post exists for. Taking every about and every non-FAQ mainEntity value across the corpus gives 163 values on 104 sites. For each one we resolved what it actually points at, read the target node's type and name, and compared that name against the names the same page gave its own Organization, WebSite, LocalBusiness or Person nodes. 107 of the 163 matched the site's own name.

The type distribution says the same thing without needing the name comparison. 73 of the 163 targets were typed Organization outright. Another 29 were a more specific organisation or person type: Service, Corporation, NewsMediaOrganization, MedicalOrganization, AccountingService, HVACBusiness, ProfessionalService, Person, and in one case an Organization that was also a BankOrCreditUnion. 102 of the 163 values, in total, pointed at an organisation or a human being. Only 16 pointed at a Thing, which is the generic type you reach for when the subject is a subject rather than a party.

Read site by site rather than value by value it is starker. On 88 of the 104 sites, every single about or mainEntity value named the site itself. On 12 of the 104, none did. The middle ground, where a site names itself on one page element and a topic on another, holds four sites. This is not a population of authors making a mixed job of a hard property. It is two populations: one using about to repeat its own identity, and a much smaller one using it as specified.

The twelve are worth naming because they show what the property looks like when it works. Bettersheets.co points about at a SoftwareApplication named Google Sheets, which is a genuine statement that the site is about somebody else's product. Etobicokerehab.com points at three Thing nodes named Chiropractic, Chiropractic treatment techniques, and Chiropractic and Manual Therapies. Trustmrr.com points about at a Dataset and mainEntity at an ItemList named Top startups by verified revenue, which is an accurate description of what that page is. Bonhomiegin.com points at a Product named Gin Bonhomie. None of those four sites gains a rich result from the markup. They are describing their pages for something that reads descriptions.

The references resolve, which is what makes this worth saying

There is an obvious explanation for the figures above that happens to be wrong, and ruling it out took a second pass over every page. 116 of the 163 values are not inline objects at all. They are bare references of the form about with a single @id key and nothing else, which is the compact way to point at a node defined elsewhere in the same document. A bare reference that points at nothing is a dead end, and if most of them dangled then the finding would simply be that the markup is broken, which is a far less interesting claim and one we have already made in another form: 117 of 612 home pages with JSON-LD carried a defect.

So we refetched all 129 sites, rebuilt each page's whole graph across every JSON-LD block on it, collected every @id the page defines, and checked each reference against that set. 115 of the 116 resolved. 113 of those resolved to a node that carried a name. Exactly one reference dangled.

That result matters more than it looks. It means the pattern is not an accident of tooling, not a half-finished migration and not a validator error. The pages are internally consistent. Somebody, or something, wired about to a real node with a real name, and the node it chose was the publisher's own. The markup is doing exactly what it was told to do. We have written before about the mechanics these references depend on, since 313 of 615 home pages gave no schema node an @id and a reference needs one at the other end; this corpus is the half that got that part right.

Two smaller patterns in the same data are worth recording because they show the property being used for things that are not subjects at all. Seven of the WebPage targets belong to one design studio whose home page declares seven mainEntity values pointing at pages named Work, Expertise, About, Latest, Awards, Contact and Clients. That is the navigation menu, declared as the primary entity the page describes. All 16 of the WebContent targets belong to a single cancer centre whose mainEntity values name promotional blocks, among them Award-Winning Excellence, Year After Year and 3X Match: Breast Cancer Awareness Month. In both cases the property is being used as a container for page furniture rather than as a statement of subject.

knowsAbout is a list, and three sites hold 178 of the 491 entries

KnowsAbout behaves differently from the other three and deserves its own count, because it is the only one of the four that is explicitly about expertise rather than subject matter. It hangs off the Organization or Person node rather than off the document, and the specification's wording is unusually hedged: a topic that is known about, suggesting possible expertise but not implying it, and it adds that no skill levels are distinguished.

39 sites declared it, carrying 491 values between them. The distribution is the point. One static analysis vendor declares 140 values on its home page alone. The top three sites hold 178 of the 491. The median across the 39 is ten. Two sites declare exactly one. A property with a median of ten and a maximum of 140 is not being used to make a claim; it is being used to enumerate a taxonomy.

The shape of the values says the same. 302 of the 491 are plain strings, which the specification permits, since knowsAbout accepts Text as well as Thing and URL. Another 142 are bare URL strings. Only 47 are objects with a name, and only 24 of those carry a sameAs, which is the property that would let an engine resolve the topic to a known entity rather than guess at a phrase. So of 491 declared areas of expertise across the corpus, 24 are tied to anything an engine could look up. That ratio is the same one we found when we counted identity references: a property that supports external resolution is mostly used without it. Entity confidence is the thing these links are supposed to buy, and a free text topic string buys very little of it.

None of this makes knowsAbout a mistake. A list of a hundred and forty topics is a reasonable thing for a developer tools company to publish, and it is more information than the 483 sites that named no subject at all. It is just not the same act as naming what one page is about, and a count that pooled the two would hide the gap this post is measuring.

Where these properties come from, and why that explains the pattern

157 sites declared isPartOf or mainEntityOfPage, the two properties that place a document rather than describe it. On 69 of those 157, the page also carried the string that a single widely used WordPress search engine optimisation plugin leaves in its output. Reading the generator meta element on the same pages adds the rest of the picture: 22 of them declared WordPress and the plugin together, 14 carried the plugin with no generator element at all, 14 ran Google Site Kit beside it, seven ran a Divi theme and five ran WPML.

That is the mechanism. These properties arrive as a side effect of installing software, not as an editorial decision, and the software emits the pattern that is cheap and safe for it to emit. A plugin running on a home page knows two facts for certain: the site has an organisation, and this document belongs to that site. It does not know what the page is about, because nothing on the page tells it in a machine readable way. So it writes isPartOf pointing at the WebSite node and about pointing at the Organization node, both of which are true, neither of which answers the question the property was defined to answer. This is the same mechanism we found when 27 of 165 article pages declared a headline the page's own h1 does not contain: a generator filling a required slot with the value it can reach.

It also explains the stratum ordering from earlier. Commercial sites run plugins and therefore declare these properties by default, with the publisher as the subject. Government sites run bespoke stacks and therefore declare nothing. Neither population is making a claim about subject matter, and only one of them looks like it is.

Google's own documentation is consistent with treating the pattern as harmless rather than useful. Its Organization structured data page, Last updated 2026-09-08 UTC, lists no required properties at all, and its recommended list runs from name, logo, sameAs and description through address to seven registration identifiers including duns, leiCode and vatID. Neither about nor knowsAbout appears anywhere on it. The one place Google requires mainEntity is the Q and A page documentation, Last updated 2026-09-08 UTC, which states that the Question for the page must be nested under the mainEntity property, which is the container use rather than the topic use. Of the 25 features in Google's structured data gallery, read on 8 October 2026 and carrying Last updated 2026-06-15 UTC, none turns about into a search feature. The properties are vocabulary rather than rich result inputs, which is precisely why they are interesting to anyone optimising for something that reads vocabulary.

What this measured, and what it did not

One page per hostname, which means a home page. A home page is the page for which the publisher genuinely is the subject, so the finding is measured on the most forgiving possible sample. An article page whose about pointed at its publisher would be a clearer error than anything counted here, and this run did not look at article pages. If the pattern is a plugin default then it will appear on interior pages too, but that is an inference from the generator counts rather than something measured, and we have not measured it.

No JavaScript was executed, so a site that injects schema from a script was recorded as carrying whatever its raw HTML carried. That is a deliberate choice and it is also a real gap: 27 of 404 home pages with no structured data in the HTML gained some when a browser ran the page, and the same could be true of these properties. Everything here is what a crawler that does not render sees, which is the population most AI crawlers belong to.

One network location, one request per URL, one date. Rate limits and bot defences remove sites unevenly, and the 215 home pages that answered 403 are not a random sample of the corpus: they skew towards sites behind the defences that also tend to run the most deliberate markup. The 612 denominator is the pages we could read, not the pages that exist.

The name comparison has a known weakness in both directions. Matching a target node's name against the page's own Organization and Person names will count a subsidiary or a parent as the site itself when the names coincide, and will miss a self reference written in a different form, such as a legal name against a trading name. We ran it case insensitively on trimmed strings and nothing more clever than that. The type distribution is the check on it: 102 of 163 targets were an organisation or person type, which is an independent route to the same conclusion and does not depend on names matching at all.

Finally, nothing here measures effect. We counted what pages declare, not what any engine does with it, and no part of this run asked an engine anything. We have been explicit about that limit elsewhere and it applies with full force to a property no documented search feature consumes. Whether naming a real subject in about changes how often a page is cited is a question about AI visibility that this measurement cannot answer, and we would rather say so than imply a mechanism we have not tested. The methodology page states the general version of that rule, and the scanner's own bot page documents the agent that made these requests.