Author schema: 95 of 111 article pages declared an author, and 8 named a brand as a person

What does author schema actually have to contain?

The vocabulary is small and it has not moved. The schema.org definition of the author property reads in full: "The author of this content or rating. Please note that author is special in that HTML 5 provides a special mechanism for indicating authorship via the rel tag." Two types are listed as expected values, Organization and Person, and nothing else. The Person type is defined as a person, alive, dead, undead or fictional, which is broad enough to cover a pen name and narrow enough to exclude a masthead.

Google's requirements are stricter than the vocabulary and they are where the interesting failures live. The Article structured data documentation, carrying Last updated 2026-09-08 UTC when it was read on 22 September 2026, opens by saying there are no required properties and that you should add the ones that apply. It then spends most of its best practice section on the author, and the instruction is unusually blunt: "In the author.name property, only specify the name of the author. Don't add any other piece of information." The things it names as other pieces of information are the publisher's name, which belongs in the publisher property, a job title, which belongs in jobTitle, an honorific, which belongs in honorificPrefix or honorificSuffix, and introductory words such as posted by.

Four more best practices sit alongside it: include every author the page presents, give each one its own author field rather than merging them into a string, add the type and a url or sameAs so the person can be told apart from everyone else with that name, and use Person for people and Organization for organizations. Read together they describe a property that is meant to identify somebody, not to label a page. That distinction is what this measurement tests, and it is a different question from the one asked here on 24 August 2026, when seven of 102 schema nodes on five captured pages named an author and six of the seven turned out to be reviewers rather than writers. That post had five pages. This one has 111, and the pattern it found survives the change of scale in a way worth recording.

How many interior pages declare themselves an Article at all?

Before the author property can be missing, something has to say the page is an article. 257 of the 563 interior pages carried no JSON-LD whatsoever, which is consistent with the finding published here on 11 September 2026 that 141 of 382 home pages carried no structured data in the raw HTML. A further 195 carried JSON-LD that contained no Article-type node. That 195 is not a defect count and it should not be read as one: the heuristic that picked these pages followed whatever the home page linked to, and on a travel board or a software company that is frequently a landing page, a product page or a section index, none of which is an article and none of which should claim to be.

111 declared an Article, a NewsArticle, a BlogPosting, a ReportageNewsArticle, a LiveBlogPosting or a Report. The distribution across the eight sectors is the least surprising result in the whole measurement and still worth stating, because it sets what the rest of the post can and cannot generalise to. 47 of the 56 news pages read were Article-typed, against 23 of 103 in software, 10 of 74 in healthcare, 8 of 80 in finance, 6 of 69 in government, 6 of 84 in education, 6 of 54 in travel and 5 of 43 in ecommerce. News is 42 percent of the denominator on its own, so this is substantially a measurement of how news organisations mark up bylines, with a long tail of company blogs behind it.

That skew also explains why the raw rate is less interesting than the failures inside it. A crawler arriving at a news article is arriving at the one page type where a byline is certain to exist in the visible text, printed under the headline, which is the part of the page a reader sees and an extractor reads as prose. The question is not whether these organisations know who wrote the piece. It is whether the assertion reaches an AI crawler in a form that is unambiguous rather than a sentence it has to guess at, which is the same gap this blog measured on 22 September 2026 when 17 of 93 SaaS pricing pages put a price in the markup while 77 showed one in visible text. Markup and prose answer different questions, and only one of them is designed to be parsed. The wider question of which markup format carries these facts was measured here on 14 September 2026, when 22 of 385 home pages carried microdata and one carried RDFa, leaving JSON-LD as effectively the only format in use.

What the 95 author values actually named

95 of the 111 carried an author property and 16 did not. The 16 are not obscure sites: worldbank.org, unesco.org, 1177.se, helios-gesundheit.de, rki.de, yle.fi, lapresse.ca, mistral.ai, newrelic.com, dropbox.com, stripe.com, visma.com, brex.com, metlife.com, principal.com and paypal.com each served a page that told a machine it was an article and then said nothing at all about who produced it. None of those 16 carried a meta name=author element either, so on those pages there is no machine-readable authorship claim of any kind.

Of the 95 that did carry the property, 89 resolved to a non-empty name and 6 did not. Typing splits 63 Person, 20 Organization and 9 with no type at all, with a small overlap where a page declares more than one author node of different types. 12 of the 111 declare more than one author, which is the case Google's guidance about separate author fields exists for. 103 of the 111 carry a datePublished, so a date is markedly more reliably present than a person.

The number that matters most for whether any of this can be used is the last one. Only 58 of the 111 pages attach a url or a sameAs to the author. On the other 53 the author is a bare name string, and a bare name string is not an identifier: it cannot be told apart from every other person with that name, and it cannot be joined to anything the engine already knows. That is precisely the property this scanner weights most heavily in its own identity scoring, described here on 25 July 2026 in the five structural signals that tell an AI who you are, where external identifiers take the largest non-gate weight because sameAs is the only property in the set whose definition is an identity statement. The same absence at organisation level was measured on 14 September 2026, when 103 of 385 pages with JSON-LD named no organization. An author with no identifier and a page with no organisation leave an engine with two names and nothing to resolve either against, which is the practical content of entity confidence as this product measures it.

The eight Person nodes that named a brand, a team or a label

63 pages typed their author as a Person. On 8 of them the name is not a person, and this is the finding worth the post. Aftenposten declares an author of type Person named Aftenposten, which is the newspaper. bhf.org.uk declares a Person named British Heart Foundation, which is the charity. redis.io declares a Person named Redis and gives it the url https://redis.io/, so the identifier confirms that the thing being described is the company. salesforce.com declares a Person named Salesforce Newsroom, which is a desk. confluent.io declares a Person named Streamhouse Working Group, which is a group of companies. mejuri.com declares a Person named Mejuri Team and kotn.com declares a Person named By Team Kotn, both of which are groups, and the second of which begins with the introductory word Google's documentation asks you to leave out. dawn.com declares a Person named ADVERTORIAL, which is not a name at all but a content label, and it repeats that string in its meta name=author element.

Each of those was read from the live page on 22 September 2026 and each is a statement the site is making about itself, so the correction in every case is the same and it is small: change Person to Organization where the author genuinely is the organisation, and name a human where a human wrote it. The type is not decoration. It is the claim about what kind of thing the name refers to, and an engine that believes it will look for a biography, a photograph and a history of other work behind a name that has none of those because it is a limited company.

Three further pages are broken rather than mistyped, and they are worth naming because the shape of the defect is one a template produces rather than a person. cmu.edu serves an author property whose value is the literal two-character string [], which is an unrendered empty array that reached production as text. derstandard.at serves an author property whose value is an empty array. kaiserpermanente.org serves an Organization whose name is the empty string. These are the same class of error found here on 21 September 2026, when 117 of 612 home pages with JSON-LD carried a defect and on seven of them not a single block parsed, and the same class as the opening hours written as one unsplittable string in 2 of 172 home pages carrying hours a machine could read. A template that emits its own placeholder is invisible to everyone except a parser.

When the author's name is on the page but not in the markup

The opposite failure is quieter and more common. 71 of the 563 interior pages carry a meta name=author element, and on 37 of them that element is the only place a machine-readable name appears, because the page either declares no Article node or declares one carrying no author property. The HTML mechanism and the structured data mechanism disagree, and a reader pointed at the structured data alone comes away with nothing.

techcrunch.com is the clearest example of a related pattern and it is not a defect. Its Article node declares an author as a single object carrying nothing but an @id, and the node that @id points at is defined elsewhere in the same document, where the name Jagmeet Singh sits. That is valid JSON-LD and it is the correct way to reference an entity once and describe it once. It is also a resolution step, and 11 of the 111 pages require it. A parser that reads the author object literally and stops will report no name on all 11, which is exactly what the first pass of this measurement did before the reference resolution was added, and it is the reason two numbers in an early draft of this post were wrong and were recomputed before publication.

The same step fails outright on two pages. frontiersin.org declares an author of @id https://www.frontiersin.org/news/author/caroline-brogan/#author and no node with that identifier appears anywhere in the document, so the reference dangles. sendbird.com declares an author of @id #identity with the same result. In both cases a person's name is almost certainly on the rendered page and the markup points at a description that was never shipped with it. usps.com and lemonade.com each declare an Organization author carrying an @id and a legalName or a type but no name property at all.

None of this says what any answer engine does with the field. It says what is there to be read, which is the only thing a fetch of the page can establish, and it is the same distinction drawn here on 8 August 2026 when review authenticity turned out not to be in the markup. If you want to see the raw bytes a crawler receives from your own pages rather than take a summary for it, what GPTBot sees fetches and shows them. The broader question of how much any of this moves a citation was examined here on 4 September 2026, when structured data came third in a 1,100 URL citation audit, behind two things that are not markup at all.

What this measurement does not show, including what Lantad does not score

The inconvenient part first. Lantad does not read an author. SCHEMA_REQUIREMENTS in core/src/schema.ts lists exactly one required property for an Article, which is headline, and the string author appears zero times in that file. Nothing in the AI Visibility Score this product sells goes up when a site fixes any of the eight Person nodes above. That was true when it was stated here on 24 August 2026 and it is still true today, and publishing a survey of a field the scanner ignores is worth more than quietly adding a signal to make the post land better. If authorship earns a place in the score it will be because there is evidence it changes what an engine does, and this measurement is not that evidence.

Nor does this post claim that any answer engine reads the author property. Google's documentation is about Google's rich results and says so. What an AI engine extracts is a separate question that a fetch of somebody else's website cannot answer, and the method this blog uses for the questions it can answer is set out at the methodology page.

Four limits on the numbers themselves. The frame is editorial rather than a random draw: 1,027 hostnames chosen to span eight sectors, so these are rates within a committed corpus and not an estimate of the web. One page per site was read, chosen by a heuristic that follows the home page's own links, so a site whose home page happens to link a section index is represented by that index rather than by its best article. No JavaScript was executed, so markup injected after load was not seen, and that cuts in one direction only: the true rates are at least as good as these. And 335 of the 1,027 sites never reached the reading pass at all, 27 because they disallow LantadBot at the root, 8 because they disallow the interior path, 97 because no link matched the heuristic and the rest because the home page did not answer, which is its own finding and was measured directly on 13 September 2026 when 79 of 115 sites refused a crawler their robots.txt allows.

What survives all of that is narrow and checkable. On the pages in this corpus that tell a machine they are articles, an author property is usually present, a resolvable identifier for that author usually is not, and on 8 of the 63 that claim a human the name belongs to a company, a team or a content label. The first is easy and was done. The second and third are the ones that decide whether the claim can be used, and they are what AI visibility measurement is for.