AI brand visibility: 412 of 1,080 home pages named their site in neither source Google reads first

Where does AI brand visibility start?

It starts with a question that sounds too simple to measure: does the page say what it is called, in a place built for saying so? Google's documentation on site names, which carries the date 2025-12-10, answers it with a list in priority order. The system "will also consider content in og:site_name, <title>, heading elements, and other text on a home page", and above all of those it puts one source: "The WebSite structured data is most important, if you want to specify a preference."

That sentence is worth reading twice, because it draws the line this post measures. Two of the five sources are declarations. A WebSite node in structured data and an og:site_name meta element both exist for no other purpose than to state a name, and a consumer reading either one is reading an answer rather than forming an opinion. The other three are content. A title element names a page and usually carries the brand somewhere inside it, along with a tagline, a separator and whatever the template appends. Heading elements and body text are prose. A machine can extract a name from all three, and it will, but it is guessing at which words are the name.

schema.org's definition of WebSite is thin on purpose: "A set of related web pages and other items typically served from a single web domain and accessible via URLs." The name property is inherited from Thing, where it means "The name of the item", and at release 30.1 dated 2026-09-16 that is the whole of it. There is no required field, no format and no validation. Which is also true of the companion type: Google's Organization documentation, last updated 8 September 2026, states plainly that "There are no required properties; instead, add the properties that apply to your organization", recommends putting the markup on "your home page, or a single page that describes your organization", and then adds the instruction this post ends on: "Use the same name and alternateName that you're using for your site name."

So the documented position is that declaring a name is optional, that the declaration is preferred over inference when it exists, and that the two places a name can be declared are expected to agree. Nothing there is a ranking promise and this post does not turn it into one. It is the resolution order a named system publishes for one narrow job, which is what entity confidence scoring rests on when there is nothing better to rest on.

What was measured, and what was not

Two requests per hostname, in the order a polite crawler makes them. First /robots.txt, parsed for groups and evaluated for the LantadBot token at the site root under RFC 9309 group selection and longest match, with a 5xx treated as disallow. Then, where the root was allowed, the home page, read as raw bytes with no JavaScript executed. 49 of the 1,419 hostnames disallowed LantadBot at the root, 38 of them only because robots.txt was returning a server error at the time, and not one robots.txt in the corpus named LantadBot in a group of its own.

From each HTML response the parser took four things: the value of every meta element keyed og:site_name or application-name, and from every script element typed application/ld+json that parsed as JSON, the name property of every node typed WebSite and of every node typed Organization or one of the 56 organisation subtypes this run recognised. Names were compared after HTML character references were decoded and whitespace collapsed, and after that the comparison is exact apart from case. That last decision is the one to argue with: 'Newegg.com' against 'Newegg Inc.' counts here as two names, and a reader may reasonably call that one name in two dresses. The counts below are therefore a floor on agreement and a ceiling on disagreement, and the examples are printed so anybody can judge them individually.

What this does not measure is worth more space than it usually gets. It does not measure whether any answer engine reads either field. Google publishes an order for its own site name feature, and that is the only first party statement this post rests on; nothing here establishes what ChatGPT, Claude, Perplexity or Gemini do with og:site_name, and this run asked none of them. It does not measure citations, mentions or what AI says about a brand, which is a different layer of the product entirely. It reads one page per site, the home page, so a site that declares its name on an about page and not on its root reads here as not declaring it, which is a real limitation given that Google's own recommendation allows either. And it says nothing about correctness: a page declaring the wrong name scores as a declaration, because nothing in the bytes can tell a parser what the right name was. The methodology page carries the rest of the settings, including the capture caps this run inherited.

412 of 1,080 home pages carried neither declared source

583 of the 1,080 pages carried an og:site_name and 376 carried a WebSite node with a name. 291 carried both. 412 carried neither, which is 38 percent of the pages a crawler was allowed to read and could read on 26 September 2026. Every one of those 412 was requested a second time the same day, independently of the first pass: 411 answered 200 with HTML again and all 411 still carried neither source, and one did not answer the same way, so the finding is not an artefact of a single fetch.

The split between the two sources is the part that surprised this run. 292 pages carried og:site_name without a WebSite node, and 85 carried a WebSite node without og:site_name, so more than three times as many sites reach for the Open Graph field as reach for the structured data field that Google says it weighs most. Open Graph is a social sharing vocabulary published at ogp.me, not a search or answer engine input, and it is almost always emitted by the same template that emits the sharing image. Its coverage here is a side effect of caring about how a link looks in a messaging app. The field Google names first is the one fewer sites supply.

Stricter counting does not rescue the picture. Widen the search from the two documented sources to four places a name can appear as a value, adding application-name and the name of any Organization node, and 313 of the 1,080 pages still carry no declared name anywhere. 89 of the 412 name an organisation without naming the site, which is a coherent thing to publish and does not answer the question Google's order asks; 66 pages in the whole corpus set application-name, a field defined for installed web applications rather than for publishers.

None of this is about markup being absent altogether. 121 of the 412 pages carry JSON-LD of some kind, so a third of them are already shipping structured data and simply do not use it to say what they are called. That is the same pattern an earlier run found from the other direction, when 103 of 385 pages with JSON-LD named no organization at all: the markup is present, the identity is not the thing it was used for.

Which kinds of site declare a name?

The corpus is stratified twice, once by what a site is built to do and once by what it is built with, and on this question the second stratification explains almost everything. Not one of the 54 Wix and Squarespace home pages was missing both sources. 1 of 36 WordPress sites was, 3 of 34 Shopify stores, and 5 of 59 news sites. At the other end, 35 of 42 Webflow sites carried neither, along with 22 of 31 documentation sites and 20 of 31 Framer sites.

That ordering is not a ranking of care. It is a ranking of whose template writes the markup. A hosted site builder ships a head section it controls, and the head section it controls includes a WebSite node and an og:site_name because the vendor decided once, for every customer, that it should. The same mechanism produced the finding that not one of 69 Wix robots.txt files named an AI crawler: the platform decides, the site owner never sees the question. A design tool that exports a page gives you complete control of the head and no defaults in it, which is why the Framer fix guide exists and why the pattern repeats across every design led stack in this corpus.

The industry strata sit in the middle and are flatter than the platform ones, which is the second half of the finding. 44 of 87 government home pages carried neither source, 46 of 95 finance, 47 of 104 education, 43 of 96 healthcare. These are the best resourced sites in the corpus, many with dedicated search teams, and they are missing a declaration that a small business on a hosted builder gets for free. SaaS came out better at 32 of 117 and news best of all at 5 of 59, the latter almost certainly because news templates have carried publisher markup since long before any AI crawler existed.

So the honest summary of the gradient is that this signal is bought, not earned. If it is present, a vendor put it there. If it is absent, nobody decided it should be absent either. That is a different kind of problem from the ones this blog usually reports, and it has a cheaper fix: four lines in a head section, once, for the whole site.

When the two declared sources disagree

291 pages declared a name twice, which creates a way to be wrong that a page declaring it once does not have. 29 of the 291 gave two different names under the comparison described above. 23 of those are cases where one string contains the other, which is mostly a legal name against a trading name: 'Frontiers' against 'Frontiers Media S.A.', 'Inserm' against 'Inserm, La science pour la santé', 'American Medical Association' against 'American Medical Association | AMA'. A human reads those as one organisation. A system applying a documented priority order reads the first one it finds and moves on, which is exactly the behaviour Google's Organization page is trying to prevent when it says to use the same name in both places.

Six are wholly different strings, and those are worth naming. On edx.org the WebSite node is named 'Build new skills. Advance your career.', a tagline rather than a name, while og:site_name reads 'edX'. On bbva.com the two read 'NEWS BBVA' and 'BBVA NOTICIAS', the same title in two languages. visitbritain.org offers 'VisitBritain.org' and 'VisitBritain Corporate'. newegg.com offers 'Newegg.com' and 'Newegg Inc.' On gitlab.com, in the contains bucket rather than this one, the WebSite node is named 'about.gitlab.com', which is a hostname where a name should be.

That last habit is common enough to count on its own. 46 of the 583 og:site_name values are the hostname itself, including Census.gov, Texas.gov, nhs.uk, Nike.com, Walmart.com, BostonGlobe.com, Trendyol.com, FRANCE.FR and, on kormany.hu, the string https://kormany.hu. A domain is an address, not a name, and a system that takes it as the site name produces answers that read like an address bar. Nothing measured here says an engine did that. It says the field that exists to prevent it was filled with the thing it exists to replace.

Set this beside what the same corpus said about external identity a day earlier. 85 of 317 organizations named a Wikipedia or Wikidata entry through sameAs, which is how a page points at an entity somebody else already resolved. That is the harder half of identity and more sites do it than declare their own name in the field Google reads first, which is a strange ordering of effort. The defect rate in the markup itself is a third, separate problem: markup can be present, be about identity, and still not parse.

What to check on your own site

The check is small and takes one view source. Open your home page, read the raw HTML rather than the rendered page, and look for two strings: og:site_name, and a node typed WebSite carrying a name. If both are there and identical, this whole subject is closed for you and nothing below applies. If one is missing, add it. If both are there and different, pick one and make the other match, which is the instruction Google's Organization page gives in its own words.

Read the bytes rather than the DOM, because on this corpus that distinction has teeth. A head section assembled by a client framework after hydration is not in the response, and 11 of 271 pages returned no words at all until a bundle ran in an earlier run. What GPTBot sees prints the raw view of one URL if you would rather not read markup by hand, and the crawlability study is where this corpus and its limits are written down.

Two things not to do. Do not put your tagline, your locale or your domain in the name field: this run found 46 hostnames and at least one slogan sitting there, and each of them is a correct field filled with the wrong kind of value. And do not treat a declared name as a ranking tactic, because nothing in Google's documentation offers one and this post measured no ranking, no citation and no answer. The claim here is narrower and harder to argue with: when a page declares its name, a consumer reads a value, and when it does not, a consumer parses a title and guesses. 412 of 1,080 pages make it a guess.

Where that matters is one layer up, in generative engine optimization and answer engine optimization, where the thing being optimised is whether an engine attributes an answer to you by the name your customers use. A page that has never stated that name in a field has delegated the decision to a title tag written for a search result and a template. Whether the delegation costs anything is not settled by this run and would need engine side observation to settle, which is what the AI Overviews guide covers on the retrieval side. What is settled is the input, and the input is free to fix. If you want these two requests run against your own pages, with the parse printed rather than summarised, that is what a scan does.