Structured data for AI answers, measured

Structured data is the schema.org markup — usually a JSON-LD block in the page head — that states what a page is and who published it, so a machine does not have to guess. Every page ranking for “structured data for AI answers” says it gets a site cited by AI. None of them counted anything. We did, across the same 75 US searches behind our citation study.

What the two engines actually say

Google, in its guide to generative AI features, last updated 10 July 2026:

Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.

However, it’s a good idea to continue using it as part of your overall SEO strategy, as it helps with being eligible for rich results.

Bing, from Fabrice Canel and Krishna Madhavan, principal product managers at Microsoft Bing, on the Bing Webmaster Blog, 20 November 2025:

Structured content, like schema-marked product pages, FAQs, and comparison tables, helps AI systems interpret and summarize your content more effectively.

Those are different sentences. Google says not required. Bing says helps interpretation. Neither says that markup gets a page cited. Most of what is written about structured data for AI makes a stronger claim than either engine has, and quotes one of them without the other.

The academic paper that named the field is worth adding. GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024) tested nine ways of making content more visible to generative engines. Every one of the nine is a change to the visible prose — adding statistics, quoting sources, citing references. The words schema, structured data and markup do not appear in the paper.

What we measured

Seventy-five US commercial SEO searches, captured by hand in September 2026, with the organic top ten and every domain the AI Overview cited recorded for each. The same corpus as the citation study.

On 14 September 2026 we then read the markup two ways:

• Pages. Every URL in the organic top ten — 546 unique — fetched and its JSON-LD parsed. 485 were read. Sixty-one refused: 53 returned 403, and 29 of those were help.shopify.com.
• Domains. Every domain an AI Overview cited plus every domain that ranked — 276 unique — fetched at the homepage with its robots.txt and llms.txt. 248 were read.

The fetcher and the JSON-LD reader are the ones behind the free audit, so this is the same check that tool makes on one site, run across every site in the corpus. We parsed JSON-LD only; a page using microdata or RDFa was counted as having markup, not as having none. A JSON-LD block that fails to parse was counted separately, because a parser — ours or Google’s — reads nothing from it.

One limit up front: the AI Overview citations in the corpus are recorded as domains, not URLs, so cited sites are compared with ranking sites by their homepages. That measures whether a site has set itself up as an entity, not what was on the particular page the overview drew from.

What ranking pages carry

Markup is the norm among pages that rank. Nine in ten have some, and eight in ten have JSON-LD. Whatever separates one ranking page from another, it is not the presence of structured data, because nearly all of them have it.

Structured data does not sort the top of the page either:

The top three carry less JSON-LD than positions four to six. Of the 44 position-one pages we read, 32 have JSON-LD and five have FAQPage. On 33 of the 75 searches, a page in the top three carries no JSON-LD at all.

FAQPage at 15% deserves a sentence. FAQPage is the schema type every page on this search recommends first. Google narrowed its rich result to government and health sites in August 2023, then stopped showing it on 7 May 2026 and removed the documentation, because “The FAQ rich result feature is no longer shown in Google Search results.” One ranking page in seven carries it; one position-one page in nine does.

Corrected 28 September 2026: the first version of this page said Google still showed the FAQ rich result to government and health sites. It had stopped on 7 May 2026, four months before this page was published.

What cited domains carry

The question the ranking pages claim to answer. Homepages, entity markup:

The domains AI Overviews cite without ranking are the least marked-up group, not the most. The gap is twelve points on 46 domains, which is suggestive and not statistically significant, so the sentence this supports is no measurable advantage, not a penalty.

Much of that first group is platforms. Remove YouTube, Reddit, LinkedIn, the app stores and Google’s own properties and the three groups sit at 85%, 91% and 90% — no difference at all. Cut either way, the sites AI Overviews cite without ranking are not more structured than the sites that rank.

Weighted by citation rather than by domain: of the 348 citations in the corpus, 236 (68%) went to a domain with JSON-LD on its homepage and 78 (22%) to a domain without. Thirty-eight of those 78 are YouTube and twelve are Reddit — the two most-cited sources in the whole corpus, neither of which ranks organically on any of the 75 searches, and neither of which serves JSON-LD on any page we could read.

Reddit is the sharper case. Its robots.txt blocks OAI-SearchBot, ChatGPT-User, Claude-User and PerplexityBot — four crawlers that fetch pages to answer live questions — and Google’s AI Overviews cite it twelve times regardless, because Googlebot is not on that list. We counted who blocks what in almost nobody has decided about AI crawlers.

The search that sells it

The search for “structured data for AI answers” returns eight pages, no ads and an AI Overview. We read the markup on all eight:

Two of the eight pages arguing that structured data gets you into AI answers have none that works.

Evertune’s page is titled “Schema vs. No Schema: Does Structured Data Matter for AI Search?” It ships two JSON-LD blocks and both fail to parse: each description string contains a raw line break, which JSON does not permit, and the first block’s @id ends in the unfilled template …/insights-on-ai/[Slug] where the article’s own slug should be. A block that does not parse yields nothing — our reader found zero nodes on the page, and Google’s Rich Results Test reports the same defect as a parsing error. We read the page on 14 September 2026; the capture that recorded the overview is from 5 September, and the fault is a template bug repeated in every description, so it very probably predates the capture. Probably is the honest word.

BrightEdge’s page, “Structured Data in the AI Search Era”, recommends JSON-LD in its own body and carries none. Its homepage carries none either.

The AI Overview on this search attributes its sentences to six sources: Stridec, Reddit, Insightland, The HOTH, Evertune and BrightEdge. Two of the six are the two pages above. A third is Reddit, quoted by the overview as the authority for the sentence “clean schema markup helps AI systems source information efficiently” — a domain with no JSON-LD on any page we could read.

None of the eight pages reports a measurement of its own. The figures they carry — “72.6% of first-page results use schema”, “20 to 30% more often in AI-generated summaries” — are attributed to third parties without a method or a link that can be checked. The one page whose title promises a test reports none.

What to actually do

Shorter than the eight guides say, and in a different order. Each item is tied to something above.

1. Set up the entity first. An Organization block with sameAs links to the profiles you control, and a WebSite block, on every page. Google states it uses structured data to “gather information about the web and the world in general”; this is the part of your markup that feeds that. It is also the part 63% of cited-only domains have and 37% do not. How many sites do all of it — one entity, an @id, the site tied to it, sameAs — is a third.
2. Point every article at that entity. Article or BlogPosting whose publisher references the Organization’s @id, and whose author is a Person or Organization node referenced the same way. An author given as a bare string is a second, disconnected entity.
3. Render the JSON-LD on the server. One GEO-tool vendor in the corpus, tryprofound.com, has an Organization block that exists only after JavaScript runs — it sits inside the framework payload and is written into the page client-side. Google renders JavaScript and sees it. A fetcher that does not, which includes the retrieval crawlers answer engines send, never does.
4. Make it valid. One raw line break inside a string and the block is ignored. Eleven of the 485 ranking pages we read — 2% — carry at least one block that fails to parse. Paste it into a validator before it ships, every time.
5. Match the visible text. Google’s guideline is one line: “Don’t mark up content that is not visible to readers of the page.” An FAQ block whose schema text differs from what a reader sees is a manual-action risk for no gain.
6. Use FAQPage where a visible FAQ exists, and expect nothing from it. Google has shown no FAQ rich result since May 2026, and its Rich Results Test no longer checks the type. It still describes the page accurately, which is the reason to keep it; check it in the Schema Markup Validator.
7. Do not count llms.txt as work done. We found a real one on 125 of the 276 domains — 45% — cited and ranking sites alike. Google’s July 2026 guide says it is fine to publish one and that “Google Search ignores them.” Keep it if another system reads it; do not mistake it for structured data.

Nothing on that list is a way to get cited. It is the way to make sure that when a system does look at the page, it reads what you meant.

What this does not show

• Not that structured data does nothing. Google says not required, Bing says it helps interpretation, and our count shows no measurable advantage in being cited. Three different sentences; we are entitled to all three and to none stronger.
• Nothing at page level about cited pages. Citations were recorded as domains, so the comparison is homepage against homepage.
• JSON-LD only. Microdata and RDFa pages were flagged as marked-up but not parsed, so 82% is a floor for “has markup”, not a ceiling.
• One moment, one vertical. Markup read on 14 September 2026 against searches captured on 5 September; 75 commercial SEO queries, US only.
• The 61 pages and 28 domains that refused us. Twenty-nine of the pages are help.shopify.com, so the Shopify searches are under-read.
• No causation, in either direction. A census with no intervention is a correlation. The version worth more is the same count in six months.

Why we ran it

Because the claim is everywhere, it is sold as a service, and it is checkable. Our GEO work starts from the entity setup in the list above — the part of structured data both engines agree has a job — and measures whether anything is citing you afterwards, rather than assuming the markup did it.

The free audit runs this exact check on your own site: what JSON-LD is there, whether it parses, whether the Organization block exists and claims any profiles. Same reader, one domain at a time.