Structured Data for AI: What to Mark Up First
By
Izzy Gregorio
·
6 minute read
How-To · Technical
Structured Data for AI: What to Mark Up First
A controlled study of 1,885 pages found adding schema moved AI citations by roughly nothing. Here is what it does do, and the order that does not waste developer time.
By Izzy Gregorio · Updated August 2026 · 10 min read
In short
Schema does not cause AI citations. A controlled Ahrefs study of 1,885 pages found no meaningful lift after adding JSON-LD. What it does is remove ambiguity about what a page is and who owns it. Start with Organization sitewide, then Article, then FAQPage on visible content. For consumer brands, product attribute markup outranks all three.
Set the expectation
Somebody finally ran the controlled test
The correlation has been quoted everywhere: pages cited by AI are far more likely to carry schema. In May 2026, Louise Linehan and Xibeijia Guan at Ahrefs tested whether that relationship was causal.
They tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, matched each against control pages from different domains with similar prior citation levels, and measured the change.
| Platform | Change in citations after adding schema |
|---|---|
| Google AI Overviews | Down 4.6%, and the only result that was statistically significant |
| Google AI Mode | Up 2.4%, not statistically significant |
| ChatGPT | Up 2.2%, not statistically significant |
Ahrefs, We Tracked 1,885 Pages Adding Schema, May 2026. Matched difference-in-differences against roughly 4,000 control pages.
“If you're already doing the rest of the SEO work well, JSON-LD isn't going to be the unlock.”
Louise Linehan, Ahrefs
A separate searchVIU experiment cited in the same report tested whether five systems read markup at the moment of retrieval. ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode all extracted visible HTML only. JSON-LD, hidden Microdata, and hidden RDFa were ignored.
Two caveats, because leaving them out would be dishonest. Every page in the Ahrefs sample already had 100 or more AI Overview citations before schema was added, so the study tests pages engines already saw. Whether schema helps a page that is currently invisible is a different question and nobody has answered it.
And both Microsoft and Google have stated publicly that structured data helps their systems understand content. The platforms say it helps. The controlled test found no citation lift. Both can be true, because they are measuring different things.
The honest framing
So what is schema actually for
It is not a ranking lever and it is not a citation lever. It is a disambiguation lever.
Structured data still powers rich results in traditional search, feeds knowledge graphs, and tells systems which entity your brand is when your name appears three slightly different ways across the web. That entity clarity is real work, and it matters more as answers get personalised.
What it will not do is manufacture authority a system was missing.
Keep it as hygiene. Do not fund it as a citation strategy.
The searchVIU finding has one consequence worth acting on immediately. If retrieval systems read visible HTML and ignore your markup, then an FAQ that exists only in schema does nothing at all.
Put the questions and answers on the page as readable text first. The markup describes them. It does not substitute for them.
The order
Four types, in implementation sequence
The work is finite. Four types cover most of the value for most businesses, and the order matters more than the completeness.
- 1
Organization, sitewide
The highest-value markup on most sites and the most frequently skipped, because it produces no visible result. It declares that your brand is a specific entity with a specific name, a canonical URL, and known alternate names.
If your business appears as three slightly different names across the web, this is where you tell the system they are one thing. Include name, alternateName, url, logo, sameAs pointing at your verified profiles, and contact details matching your listings exactly.
- 2
Article, on every content page
Declares authorship, publication date, and modification date. The author must be a real named person with a linked bio page, not a company. This is the single most common gap on otherwise well-built sites, and experience and expertise signals carry weight in what a system is willing to state as fact.
- 3
FAQPage, matching visible content verbatim
Question-shaped markup on question-shaped content. The questions in your schema must match the questions on the page exactly. Given that retrieval reads visible HTML, the visible FAQ is doing the work and the markup is describing it. Schema describing content that is not visible is a policy violation, not a shortcut.
- 4
HowTo, on process content only
Do not add it to a post that is not actually a process. Over-marking is a real failure mode and it teaches systems to trust your markup less.
Translate it
Where the highest-value markup is, by business
| If you are a | The markup that matters most | Why it outranks the general list |
|---|---|---|
| Consumer or CPG brand | Product, with full attribute coverage | Google's guidance to CPG brands is direct: product data is the new packaging, and untagged attributes are invisible to agents |
| Local service business | LocalBusiness, plus exact NAP consistency | An assistant naming one to three local businesses needs unambiguous location and service data |
| Professional services firm | Service and Person, plus Organization | Credentials and specialisms need to be declared in text, not implied by page design |
| Publisher or content brand | Article and Author, with real bios | Authorship signals decide what a system is willing to attribute |
| Nonprofit or ministry | Organization, Event, and NonprofitType | Gathering times and service areas are the most-asked and least-marked-up facts |
The consumer brand row is the one exception to everything above, and it is worth pausing on. AI traffic to US retail sites grew 1,324 percent between October 2024 and May 2026 according to Semrush and Adobe data, and Adobe Analytics found AI-referred shoppers converted 42 percent better than traditional traffic in the first quarter of 2026.
If an agent is assembling a shortlist from product attributes, unmarked attributes are not a missed optimization. They are an absence.
Do the work
Five steps, in order
- 1
Audit what exists. Run your key pages through Google's Rich Results Test and the Schema Markup Validator. Many sites already have partial schema injected by a theme or plugin, and it is frequently wrong.
- 2
Implement Organization sitewide. One block, one time, in the site header or footer template. Confirm every value matches your Google Business Profile and directory listings character for character.
- 3
Add Article to content pages. Include a real author with a linked bio page. If no bio page exists, build one before adding the markup. Author schema pointing at nothing is worse than no author schema.
- 4
Add FAQPage where FAQs are visible. Match the questions verbatim. If you edit a question on the page, edit it in the schema the same day.
- 5
Validate, then re-validate on a schedule. Schema breaks silently during site updates. Add it to a quarterly technical check, because nothing will notify you when a template change drops the block.
The common failures
Four ways this goes wrong
Marking up content that is not visible. Schema must describe what a human sees on the page. Invisible-content markup is a policy violation and carries penalty risk, and given that retrieval reads visible HTML, it also does nothing.
Author schema pointing at a bio page that does not exist. Extremely common, and it undermines the exact signal it was added to send. Build the bio page first.
Stale dateModified. An unchanged modified date on updated content wastes the freshness signal, and freshness is weighted by the engines that retrieve live.
Plugin-injected schema nobody has read. Themes and SEO plugins add markup automatically, and it is frequently generic, duplicated, or contradicts hand-written blocks on the same page. Read what is actually on your pages before adding more.
One question worth sitting with. Open your highest-value page and view its structured data right now.
If you cannot say who it declares as the author, when it says the content was last modified, and what entity it says your business is, then the page is currently telling a system something nobody in your company has ever read.
Common questions
Structured data for AI, answered
Does schema markup actually improve AI citations?
No measurable evidence says it does. Ahrefs tracked 1,885 pages that added JSON-LD against roughly 4,000 control pages and found no statistically significant citation uplift on Google AI Mode or ChatGPT, and a small significant decline on AI Overviews. Treat schema as a disambiguation lever rather than a citation lever.
Do AI engines read schema markup at all?
Not at the moment of retrieval, according to a searchVIU experiment cited in the Ahrefs report. ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode all extracted only visible HTML when fetching pages directly, ignoring JSON-LD, Microdata, and RDFa. Schema may still play a role during indexing and classification, which the test did not measure.
Should I remove my schema then?
No. Structured data still powers rich results in traditional search, feeds knowledge graphs, and helps systems resolve which entity your brand is. Keep it as hygiene. The narrower point is that it should not be funded as an AI citation strategy or sold as one.
What schema types should I add first?
Organization sitewide, then Article on content pages, then FAQPage where visible FAQs exist, then HowTo only on genuine process content. For consumer brands, product attribute markup takes priority over all of these, since agents assembling shortlists read product data directly.
Can I add schema for content that is not on the page?
No. Structured data must describe content visible to a human visitor. Marking up hidden or non-existent content violates published guidelines and carries penalty risk. It also achieves nothing at retrieval, since the systems read the visible page rather than the markup.
If schema is not the lever, what is?
Earned media, entity clarity, and content extractability. Roughly 82 percent of AI citations come from earned media per Muck Rack's December 2025 analysis, and research presented at KDD 2024 found adding statistics produced the largest citation lift of any tactic tested. Schema makes your pages easier to classify. It does not make systems reach them.
Get the data
The Citation Gap Report 2026
Where structured data sits relative to the levers that do move citations, with the data behind each one. So you can size this work honestly before you assign it to a developer, and spend the rest where it counts.