Adding schema is not a demonstrated shortcut to more AI citations. An Ahrefs observational study found no clear citation uplift on already heavily cited pages after JSON-LD was added. That narrower result is useful when deciding what to prioritise.
The honest answer first
In May 2026, Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, against 4,000 matched pages that didn’t. The result: ChatGPT citations up 2.2%, Google AI Mode up 2.4%, both statistically indistinguishable from zero — and Google AI Overviews down 4.6% relative to matched controls. Ahrefs cautions that the small decline cannot confidently be attributed to schema. The study pooled schema types, covered a 30-day window and could not fully separate concurrent changes. Google’s own guidance, published the same month, says it plainly: “there’s no special schema.org markup you need to add” for generative AI features.
So why does this deserve 1,200 words? Because in the same Ahrefs dataset, 53% of AI-cited pages carry schema and cited pages were nearly three times more likely to have it than uncited ones. That is a correlation, not proof of a causal mechanism. The study does not establish schema’s effect on previously unseen pages, or whether entity identification happens before or after a citation.
What schema actually does for a machine
Treat JSON-LD as a server-rendered entity ledger: who you are, who wrote this, what it costs, when it last changed. How AI retrieval systems use that markup varies and is not fully documented. Mark Williams-Cook built a fake company with nonsense JSON-LD in May 2026 and the tested systems still returned the address. That small experiment shows information recovery despite invalid markup; it does not establish how every AI system parses pages or selects sources.
Microsoft has also discussed schema’s role in understanding content. At SMX Munich in March 2025, Microsoft’s Fabrice Canel said schema helps Microsoft’s LLMs and Copilot understand content, and pointed site owners at IndexNow to push freshness signals as they publish. A reported conference remark does not establish an internal retrieval mechanism or a guaranteed citation benefit.
Some rich-result features have been retired
Google restricted FAQ rich results to “well-known, authoritative government and health websites” and made HowTo desktop-only on 8 August 2023; a follow-up update killed HowTo entirely five weeks later. In June 2025 Google retired seven more structured-data features — Book Actions, Course Info, Claim Review, Estimated Salary, Learning Video, Special Announcement, Vehicle Listing — because they “aren’t commonly used in Search” and provide “no longer […] significant additional value” (Google, 12 June 2025). Then, on 7 May 2026, FAQ rich results disappeared for everyone; Google pulled the documentation five weeks later and the old FAQPage URL now redirects to the changelog.
Through all of it, Google’s line on unused markup has stayed consistent: it “does not cause problems for Search, but also has no visible effects.” Stop writing schema chasing snippets that no longer exist.
Make the markup accessible
Markup is useful only to a system that can access it. Vercel’s December 2024 traffic analysis reported no JavaScript rendering by several AI crawlers in its observed dataset. That is evidence from a particular period and set of sites, not a permanent specification for every current crawler or browser agent.
Serving essential JSON-LD in the initial HTML avoids depending on client-side execution. Google can process JavaScript-generated structured data, but other systems differ; check the initial response and the relevant platform’s rendered view where available. Keep the markup accurate and consistent with the visible page. Google’s JavaScript SEO guidance.
The four blocks that still earn their place
Organization: the anchor
Google recommends the home page or a page describing the organisation; it says repeating the information on every page is unnecessary, not prohibited. Include applicable, accurate properties such as @id, name, legalName, url, a logo at least 112×112px, contactPoint, foundingDate, and — for UK entities — an identifier that actually disambiguates you. Google’s Organization guidance (updated 15 April 2026) lists iso6523Code, leiCode and vatID, and states ISO 6523 is preferred as “less ambiguous.” There are no universally required Organization properties. Use genuine identifiers that apply to the business; do not invent one to complete a checklist.
{
"@type": "Organization",
"@id": "https://example.co.uk/#organization",
"name": "Example Ltd",
"legalName": "Example Limited",
"url": "https://example.co.uk",
"logo": "https://example.co.uk/logo.png",
"sameAs": ["https://www.linkedin.com/company/example"],
"vatID": "GB123456789"
}Article and Person: bylines that resolve
Google’s Article documentation (updated 10 December 2025) has no strictly required fields, but its author rules are specific: list every author actually shown on the page, type them as Person or Organization (never Thing), keep author.name to the name only — no “Dr”, no job title — and link to a bio page via author.url or sameAs. Use ISO 8601 dates with timezone information, because Google defaults to Googlebot’s timezone if you omit it. Only update dateModified when the content actually changed; a build timestamp is not evidence of a substantive update.
Product and Offer: check the relevant shopping integration
Product markup and merchant feeds serve particular shopping experiences. Whether they are relevant depends on what the business sells. Google’s Shopping Graph held more than 50 billion product listings in May 2025, with over 2 billion refreshed every hour, feeding AI Mode shopping results (Google, 20 May 2025). But OpenAI’s ChatGPT product-feed spec doesn’t mention schema.org at all — it wants a JSONL, CSV, TSV or TXT feed with nine required fields (item_id, title, description, url, brand, seller_name, image_url, availability, price). If those integrations are relevant to your store, maintain their required data and validate each separately. A UK service business does not automatically need a product feed.
FAQPage: keep it only where the Q&A genuinely exists
There have been no FAQ rich results anywhere since 7 May 2026, and Google says leaving old markup in place causes no harm. What’s left of FAQPage’s value is machine-readable Q&A content — worth keeping on a genuine FAQ page, never worth injecting by template across pages that don’t actually contain questions and answers.
sameAs and @id: link, don’t list
sameAs should point to profiles that genuinely describe the entity — Google specifies “a page on another website with additional information about your organization,” such as a social profile or review site, not a random directory listing. Use one @graph with stable @id values (#organization, #website, and an article’s publisher pointing back at the organization’s @id). A doubled https:// inside a URL field is an error worth correcting. Validate the affected property and consumer behaviour; a malformed URL does not by itself establish that the whole structured-data array is ignored.
What to do
- Prefer essential JSON-LD in the initial HTML where practical; also check the rendered view for platforms that process JavaScript.
- Describe the organisation accurately, using applicable properties and identifiers rather than filling every optional field.
- List visible authors as
PersonorOrganization, as appropriate, with the name and a relevant identifying URL. - Only change
dateModifiedwhen the page genuinely changed. - Drop FAQPage from templates; keep it only on pages with real Q&A content.
- Run both Rich Results Test and Schema Markup Validator — they check different things.
What we changed in our own scoring
We had a “30–40% higher AI citation rate” line in our own schema copy. It wasn’t sourced, so we cut it. The Ahrefs result does not justify replacing it with an equally universal claim that schema can never help. FAQPage is no longer a warning in our scoring — it’s informative, reflecting the 7 May 2026 deprecation — and Organization completeness (logo, sameAs, identifiers) is reported alongside the scored checks, but is informative and contributes no points. Schema sits at 15% of an AIEO score because that is our chosen rubric. The weight is not a measured contribution to rankings or citations.
Want to see how your own structured data reads to an AI crawler? Run the free scan.