Schema Markup and AI Citations: What Google, Bing and 1,885 Pages Show

Adding schema markup to a page does not measurably raise how often AI answers cite it. That is the result of the largest controlled test published so far, in which Ahrefs tracked 1,885 pages that added JSON-LD and compared them with matched pages that did not. Google says no special structured data is needed for AI Overviews. Bing says schema helps its AI understand content. Both statements can be true, and neither one is a citation lever.

The direct answer, for anyone deciding where to spend engineering hours: schema markup is worth implementing for the rich results and entity clarity it still earns in classic search, not for AI citations. Build Organization, Article, BreadcrumbList and, if you sell things, Product. Skip anything sold as AI-specific schema, and do not start new FAQ or HowTo work for Google.

What Google says about schema markup and AI Overviews

Google has put its position in writing, in the documentation page titled AI features and your website. The page was last updated in December 2025. It is unusually blunt for Google documentation.

The documentation sentence that settles eligibility

The page states that there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary". It then goes further on markup specifically. "You don't need to create new machine readable files, AI text files, or markup to appear in these features," it says. It adds that "there's also no special schema.org structured data that you need to add".

The only structured data advice on the page is a hygiene rule. Make sure your structured data matches the visible text on the page. That is a warning against markup that claims things the page does not show, not an invitation to add more of it.

The eligibility rule is equally plain. A page qualifies as a supporting link in AI Overviews or AI Mode if it is indexed and eligible to appear in Search with a snippet. Nothing in that sentence depends on JSON-LD.

Google's general introduction to structured data still describes the markup as "explicit clues about the meaning of a page." It recommends JSON-LD as the easiest format to maintain, while saying all three formats are equally fine when valid. So Google uses structured data to understand pages. It simply declines to say that understanding becomes AI citations.

What Google keeps removing

The more telling evidence is what Google has been switching off. Each removal shrinks the set of schema types that produce anything visible in its results.

In August 2023, Google restricted FAQ rich results to well-known, authoritative government and health websites. A month later it stopped showing HowTo rich results on desktop, which completed that type's deprecation. In June 2025 its simplifying search results post phased out seven more: Book Actions, Course Info, Claim Review, Estimated Salary, Learning Video, Special Announcement and Vehicle Listing. Book Actions was later reprieved.

November 2025 added practice problem and Dataset markup to the list. Then the documentation changelog recorded that the FAQ rich result no longer appears in Google Search as of 7 May 2026. Google has repeatedly said these changes do not affect ranking, and that unused markup causes no problems.

What Bing says about structured data for AI answers

Microsoft is the one major search operator that recommends schema in the context of AI answers. Read its wording closely, because it is narrower than the way it gets quoted.

The October 2025 Microsoft guidance

On 8 October 2025, Krishna Madhavan, a principal product manager at Microsoft Bing, published Optimizing Your Content for Inclusion in AI Search Answers. One section addresses schema directly. "Schema is a type of code that helps search engines and AI systems understand your content," it says. It notes schema can label content as "a product, review, FAQ, or event."

The same post spends far more words on other things. Assistants like Copilot "break content down" into smaller pieces that are "ranked and assembled into answers". It warns against hiding answers in tabs or expandable menus, putting key facts only in images, and relying on PDFs. The checklist line reads "Use schema, clear headings, and modular layouts."

Notice what is absent. The post never says schema raises the chance of being cited, and it never quantifies anything. It is a guide to being parsed well, with schema as one item among several.

The 2025 conference remark

The other Bing statement in circulation is older. Search Engine Land reported in March 2025 that Fabrice Canel, then a principal product manager at Bing, said on stage at SMX Munich that schema markup helps Microsoft's LLMs understand content.

Mark Williams-Cook, who has examined how that remark spread, points out that Canel's written follow-up was about something else. "Gen AIs value fresh content in particular," Canel wrote, recommending IndexNow to push updates. That is a freshness point. It became, through repetition, a claim that JSON-LD wins AI citations.

The right-hand column is the useful one. Every confident claim that schema drives AI citations has to step outside what one of these sources supports.

What the controlled schema markup tests show

Official statements tell you what a company is willing to say. Tests tell you what happens. Three published pieces of work are worth knowing, and they differ sharply in rigour.

Ahrefs: 1,885 pages, matched controls, no lift

Ahrefs published the strongest test on 11 May 2026, written by Louise Linehan with analysis by Xibeijia Guan. They found 1,885 pages that first added JSON-LD between August 2025 and March 2026. Each was matched to control pages from other domains with similar citation levels that never added schema.

They then counted citations 30 days before and 30 days after the change, across Google AI Overviews, AI Mode and ChatGPT. Four tests were run, including a matched difference-in-differences that strips out platform-wide trends. That is the method Ahrefs says it trusts most.

The results were flat or negative. AI Mode showed +2.4% and ChatGPT +2.2%, both statistically indistinguishable from zero. AI Overviews showed a 4.6% decline relative to controls, the only significant movement. Both groups were already falling, so Ahrefs calls the decline real but unexplained rather than caused by schema.

One detail is worth keeping for the next time someone shows you a before-and-after chart. Raw AI Mode citations for the treated pages rose 43%. The controls rose almost as much. Without the control group, that 43% would have been the headline.

Search Atlas: domain coverage against LLM visibility

Search Atlas published The Limits of Schema Markup for AI Search on 14 December 2025, by founder Manick Bhan. It grouped domains by the share of their sampled pages carrying schema, from none to full, and compared those bands with visibility in OpenAI, Gemini and Perplexity answers.

Its conclusion is that "domains with complete schema coverage perform no better than those with minimal or no schema." The study is weaker than it sounds. It publishes no sample size, no collection window and no significance test, and it measured presence of schema rather than type or quality. Search Atlas also sells an LLM visibility product. Treat it as a directional observation that agrees with Ahrefs, not as independent proof.

The duck test: models read JSON-LD as text

Mark Williams-Cook ran the cleverest small experiment, published in May 2026. He built a page for a fictional duck T-shirt company and put its address only inside a JSON-LD block. The block was deliberately invalid, with a non-existent context URL, a made-up type called MallardEnterprise and properties that do not exist in schema.org.

Both ChatGPT and Perplexity returned the fake address when asked. A real schema parser would have rejected that block. His reading is that the models "were not parsing it as schema," and the markup was "simply more text on the page, lightly garnished with curly braces".

He is careful about the limit of his own result. A system that consulted schema and fell back to text extraction would give the same answer. The test proves only that "a model returned a fact from the schema" is not evidence that schema was used as schema.

Why the schema correlation fooled the industry

If the tests are this clear, why do so many guides still claim schema drives AI citations? Because the correlation is real, and it is large.

Before running its controlled test, Ahrefs analysed 6 million URLs. It found that AI-cited pages were almost three times more likely to carry JSON-LD than pages that were not cited. A figure like that, lifted out of context, reads like a ranking factor.

It is better explained as a fingerprint. Sites that ship clean JSON-LD tend to be the same sites with technical SEO staff, maintained templates, authoritative content and strong links. Ahrefs says this directly. The schema is a marker of an organisation that does the work, not the work itself.

This is the same trap our analysis of whether ChatGPT citations track Google rankings runs into. Anything that travels with quality will correlate with being cited. Only a controlled design separates the passenger from the driver, and the controlled design here says passenger.

The wider evidence on which tactics actually move AI visibility, and which are restatements of a single vendor study, is covered in the review of GEO claims against published evidence. Schema is one row in that ledger. This post is the long version of that row.

How AI systems actually read your structured data

The test results make more sense once you separate three different ways an AI answer can encounter your page. Schema matters differently, or not at all, in each.

At training time, large text corpora are usually extracted from raw HTML with tools that keep visible prose. Williams-Cook notes that FineWeb, a 15 trillion token dataset built from 96 Common Crawl snapshots, uses trafilatura for extraction. His argument is that script blocks, where JSON-LD lives, are stripped before a model ever sees them. That is an inference about one public dataset, not a disclosure from any lab.

At answer time, through a live fetch, the duck test suggests models see the page source as text. Any fact in your JSON-LD may surface, but no better than the same fact in a paragraph. The Microsoft guidance points the same way from the other side. Its advice is overwhelmingly about visible structure: headings, short sections and answers not hidden in tabs.

At answer time, through a search index, schema can genuinely matter. Google and Bing both parse structured data into their indexes, and AI Overviews draw on Google's index. That is why Google says structured data helps it understand pages. It is also why the effect, if any, shows up as better understanding rather than more citations.

The practical consequence is simple. If a fact matters to a buyer, put it in visible HTML first. Markup can repeat it. Markup should never be the only place it exists, which is also Google's own rule about not marking up invisible information.

Schema markup worth the engineering time in 2026

Here is where we land. We would build a short list of schema types properly and stop. The case for each rests on classic search features and entity clarity, not on AI citations.

This page practises what it preaches. It carries Organization, WebSite, Person, BreadcrumbList, Article and FAQPage markup, generated once by a template. We keep FAQPage because it costs nothing at that point, not because we expect it to win a citation.

The ordering matters as much as the list. If your team has two weeks of SEO engineering, spend it on crawlability and rendering before markup. A page whose key answer sits behind a JavaScript tab is invisible to several AI crawlers regardless of its JSON-LD. Our breakdown of which AI crawlers to allow in robots.txt is a better first ticket than any schema work.

Measure the outcome the way Ahrefs did, at small scale. Add markup to a handful of pages, leave a comparable handful alone, and compare citations after 30 days or more. The tooling for tracking that is laid out in how to build an AI visibility dashboard.

Where this argument is weakest

The case against schema as a citation lever rests mostly on one study. Here are its gaps, stated as plainly as we state everyone else's.

The Ahrefs sample was already famous. Every page had more than 100 AI Overview citations in February 2025 before any schema was added. Pages that AI systems already cite heavily may have no headroom left. Schema could still help a page that AI has not yet noticed, and this test cannot see that case. For a new domain like ours, that is precisely the case that matters.

The window was 30 days. Indexes refresh on their own schedules. A slower effect over 60 or 90 days would be invisible here, and Ahrefs lists this limitation itself.

Types were pooled. Article, FAQ, Product, HowTo and Organization were analysed together. A real effect from one type could be diluted by the rest. Product markup in shopping queries is the most plausible place for that to hide.

Bing may simply be telling the truth

No published controlled test covers Copilot or Bing's AI answers. Microsoft says schema helps its AI understand content, and Bing's index does parse it. It is entirely possible that schema helps in Copilot and not in Google, and the evidence above would not detect that. If Bing traffic matters to your buyers, that uncertainty is a reason to keep the core types, which you should be doing anyway.

Last, every test here comes from a company that sells SEO software, and none is peer reviewed. They agree with each other and with Google's documentation, which counts for something. It does not make them independent replications.

Frequently asked questions

Where to start with schema markup this week

Open your site's template code and list every schema type it emits. Strike anything Google no longer displays, then check that Organization, Article and BreadcrumbList are present and that every value they carry also appears in visible text. That audit takes an afternoon.

If anyone on your team, or any agency you pay, has a line item called AI schema or structured data for LLMs, ask one question. What controlled test showed it worked? Ahrefs has published a design you can copy with 10 pages, and a vendor who cannot name a comparison group has not measured anything.

Related reading