We Asked 71 Pages an AI Cited What They Declare About Themselves. Fewer Than Half Say What They Are.

Every checklist in this category tells you to mark up your pages. Almost nobody has asked the pages that AI answers actually cite what they declare. We did: 93% carry some JSON-LD, 41% carry a type that says what the page is, and three carry nothing at all.

The population is the interesting part. These are not pages we picked as good examples. Every URL here was returned as a source by an AI answer in a measurement we had already run and frozen, so the engines chose the sample.

How this was measured: 71 URLs cited as sources across 9 category questions in 4 markets, each fetched once with a plain HTTP client on 18 August 2026, with a second read for anything that failed. Every JSON-LD block was parsed and its @type values collected. Pages that answered with almost no body text are counted as shells rather than as pages that declare nothing, and pages that returned a bot challenge or no answer are kept out of every count. The full rows are in our measurement register.

Disclosure: EchoWi sells AI visibility measurement, and a finding that structured data is not the differentiator in this set is a finding against a thing many tools in our category score you on. The URLs, the buckets, the date and the method are all here, and the population came out of our own frozen measurements rather than a list we assembled, so anyone can repeat this against us.

The short version

1. 93% of readable cited pages carry some JSON-LD. Publisher boilerplate is near universal: Organization, WebPage, BreadcrumbList, Article.
2. 41% carry a type that says what the page is. A comparison, a review, a product, a set of questions, a piece of software. The rest declare only who published them.
3. FAQPage is the type these pages actually carry, on 17 of 46. ItemList appears on 6 and Review on 3.
4. Three cited pages carry no structured data of any kind, and they are not empty: 29,724, 12,429 and 6,737 characters of visible text.
5. Every unreadable page in the set is a platform. 17 YouTube URLs, two Reddit and one Facebook answer with almost no body text, which is a fact about where those platforms keep their content.

What was counted, and what was refused

The unit is one URL that an AI answer named as a source. Nine questions produced them, across accounting, payroll, CRM, travel credit cards and headphones, in the United States, Spain, France and Germany.

For each URL we parsed every <script type="application/ld+json"> block and collected the @type values anywhere inside, including nested ones. Then we split the types into two groups, and the split is the whole point of the study.

Publisher types say who put the page there: Organization, WebSite, Person, BreadcrumbList, ImageObject, WebPage, Article. Every content management system emits these, and their presence tells a machine nothing about what it is looking at.

Type-of-page types say what the thing is: ItemList for a comparison, Review for a review, Product for a product, SoftwareApplication for software, FAQPage for a set of questions and answers. These are the ones a retrieval layer could use to decide that a page answers a comparison question.

Three states were kept apart rather than merged, because merging them is the most common way a sweep like this reports a number that means something else:

• Read, meaning the page answered with a body a plain client can read. Only these are counted.
• Shell, meaning the page answered with under 500 characters of visible text. That is a fact about where the page keeps its content, not about whether it has any.
• Blocked or no answer, meaning we never saw the page. Four returned a bot challenge and one never answered.

The result

And the breakdown by type is where the standing advice and the evidence part company:

The advice everyone gives about comparison content is to mark it up as an ItemList, because that is the type that declares “this page is an ordered set of N things”. Among the pages an AI actually cited for these nine questions, ItemList appears six times. FAQPage appears nearly three times as often.

Some of the pages carrying the strongest markup are exactly who you would expect: thepointsguy.com and lesnumeriques.com each carry FAQPage, ItemList, Product and Review together. But bankrate.com, creditcards.com and nytimes.com sit in the same answers with two types apiece, and frandroid.com and soundcore.com with one.

The three that declare nothing, and why they matter more than the rest

bakedwith.com, zoho.com and capitalone.com were each cited by an AI answer and carry no JSON-LD, no microdata and no RDFa. Not a stripped-down page, not a redirect: 29,724, 12,429 and 6,737 characters of readable text respectively.

That is the observation that constrains what anyone can claim here. If structured data were a requirement for citation in this category, these pages could not be in the set. They are, so it is not.

The honest reading of that is narrow, and it is worth stating precisely because the wide reading is so tempting. This does not say markup is useless. It does not say adding schema will not help you. It says one thing: you cannot look at a citation in this category and infer that the page was marked up, because 7% of them were not marked up at all and 59% did not declare what they were.

Every shell in the set is a platform

Twenty of the 71 URLs answered with almost nothing a plain client could read. All twenty are platforms: 17 YouTube watch pages at 254 visible characters each, two Reddit threads at five characters, and one Facebook page at 44.

We have measured this before on a different population, and the finding replicates exactly: these platforms keep the material in the document head and in their own rendering layer, not in the body. A YouTube page carries its title, description and a full VideoObject in JSON-LD while its body says almost nothing.

The practical consequence is not about YouTube. It is that any sweep counting “pages with no visible content” will report platform URLs as empty pages unless it separates them, and platform URLs are a large share of what AI answers cite. In this set they are 28% of everything the engines named.

The comparison group, added

The section above said this study could not tell you what uncited pages carry, because every URL in it had already been cited. That gap is now filled with the cheapest possible control: the same nine questions put to Google’s organic results, and every ranking URL the AI answer did not cite, fetched with the same client and parsed with the same code. 103 of them, 71 readable.

The direction was written down before the first call. If the controls came in more than 15 points below the cited side on the types that say what a page is, then markup would separate the two groups in this sample, the association other people have reported would replicate on our own population, and that is what we would be publishing.

The gap on citable types is 10 points, and the threshold was 15. So the headline does not flip. What is true, and worth saying without inflating it, is that all three rows point the same way: cited pages carry somewhat more markup than uncited pages that rank for the same question, by margins too small to call markup the thing that separates them.

The type-level detail is where it gets more specific than the totals:

FAQPage appears the same number of times on both sides, which over different denominators means it is proportionally more common among the cited. ItemList is three times as common among the cited. And Product runs the other way, more common among pages that ranked and were not cited, which is the row that stops this from being a story about markup in general.

One more difference is not about schema at all: 12 of the 103 control URLs answered our fetch with a bot challenge, against 4 of the 71 cited ones. That is 12% against 6%, on small numbers, and it points at something more plausible than markup as a mechanism. It is also not evidence about AI crawlers, because a page that challenges the HTTP client used for this sweep is not necessarily challenging GPTBot, and we did not test that here.

What we would take from this if we were buying

Do not buy a score that is mostly a schema audit. Several tools in this category will grade your page on markup coverage. On this evidence that grade is measuring something real about your page and something weak about your chances of being cited, because 59% of the pages that were cited do not carry the types the grade rewards.

If you are going to add one type, add the one these pages carry. FAQPage is on 17 of 46, and it has an obvious mechanism behind it: a question with a self-contained answer is the unit a retrieval layer can lift whole. That is the same reason our own corpus puts a FAQPage on nearly every page, which is a practice we adopted before this measurement rather than because of it.

Do not conclude that markup does not matter. This study cannot support that, and the reason is in the next section.

What this does not show

• The population is selected on the outcome. Every URL here was cited. That makes the study cheap and conclusive for killing one claim, that cited pages in this category are heavily marked up, and useless for its converse. A control group has since been added above, from the same nine questions, and it narrows the gap rather than closing the question: pages that rank and are not cited carry markup at a similar rate. The study with a much larger comparison is AirOps’, which reported 61% of ChatGPT-cited pages carrying three or more schema types against 25% of uncited first-page results, across more than 12,000 URLs. Ours is a smaller sample asking a narrower question on a population the engines picked, and the two are not in conflict: theirs finds an association across a much larger set, and the controlled test Ahrefs ran on 1,885 pages found adding schema did not significantly raise citations. Association, no measured lift, and no common markup among the cited pages we could reach: three different measurements pointing the same way about how much weight to put on a markup score.
• One fetch, one client. A page that builds its JSON-LD in the browser would look bare to us. That would push the real numbers up, not down, so the 41% is a floor rather than an estimate.
• Nine questions, four markets, one day. Five product categories and one snapshot. Another set of questions would produce another set of publishers.
• Nothing here is causal. We can say what the cited pages declare. We cannot say that declaring it is why they were cited, and we did not test any intervention.

Common Questions About Schema and AI Citation

Does structured data help a page get cited by an AI answer?

This study cannot answer that, and it is worth being blunt about why: every page in it was already cited, so there is no uncited comparison group. What it can answer is the narrower question people usually mean, which is whether the pages winning citations in a category are the heavily marked up ones. In these nine questions they are not: 41% declare what the page is and 7% carry no structured data at all.

Which schema type do cited pages actually carry?

FAQPage, on 17 of 46 readable pages, ahead of ItemList on 6, Product on 6, Review on 3 and SoftwareApplication on 2. That ordering runs against the usual advice for comparison content, which leads with ItemList.

Why are YouTube pages counted separately?

Because a YouTube watch page answers a plain HTTP client with about 254 characters of visible body text while carrying its title, description and a VideoObject in the document head. Counting it as a page that declares nothing would be a statement about YouTube that is false. 17 of the 20 shells in this set are YouTube.

How can a page with no schema at all be cited?

We do not know, and that is the correct answer rather than a hedge. Three pages in the set carry none and are long, ordinary, readable articles. Whatever the retrieval layer used to select them, it was not their markup, because there was none to read.

How would I run this on my own category?

Take the answers you already measure, keep the source URLs, fetch each one and parse its JSON-LD. The only two decisions that matter are separating pages you could not read from pages that declare nothing, and separating types that say who published the page from types that say what it is. Both of those are where a sweep like this usually goes wrong.

Does this change what EchoWi recommends?

It narrows one recommendation. We still put a self-contained answer and a FAQPage on our own pages, and our verified tool catalogue still carries an ItemList describing exactly what it lists. What we would no longer say is that markup coverage predicts citation, because in the only sample we have of pages that were actually cited, it does not separate them.