Schema markup for AI search: what the evidence says

Start with the study that should temper everyone's expectations. In May 2026, Ahrefs published a controlled test across 1,885 pages that examined what happens to AI citations when JSON-LD structured data is added. The result was blunt. For ChatGPT and Google AI Mode, adding schema produced no measurable citation lift at all: the movement in the data was indistinguishable from noise. For Google AI Overviews, the study found something worse than nothing, a statistically significant decline in citations after schema was added.

That decline deserves care rather than panic. Nobody has proposed a convincing mechanism by which valid structured data would actively suppress citations, and the study's own framing points to a more likely explanation: the pages tested were already heavily cited before the experiment began. Pages at the top of the citation distribution have little headroom, and any intervention measured against them will tend to show regression toward the mean. The honest takeaway is narrower than the headline: on pages that AI engines already cite, adding schema did not add citations, and in one surface the trend line went the wrong way.

What the Ahrefs study does not tell you is whether schema matters at the other end of the distribution, on new pages or unknown domains where an engine is parsing your content for the first time and deciding what it is about. That gap in the evidence is where the rest of this playbook lives, because the strongest claim anyone can honestly make for schema in 2026 is about parsing and disambiguation, and that claim is plausible precisely where the Ahrefs data has nothing to say.

Now the other side of the ledger. SE Ranking analyzed the pages AI engines actually cite and found that roughly 71 percent of pages cited by ChatGPT carry structured data, and about 65 percent of pages cited by Google AI Mode do. If you looked at that number alone, you would conclude schema is close to a prerequisite for citation. Seven in ten cited pages have it.

The number is real, and the inference is where people go wrong. This is a correlation, and SE Ranking presents it as one. Sites that implement schema correctly tend to be professionally operated: they have working templates, fast pages, clean information architecture, an SEO function, and content someone maintains. Every one of those qualities plausibly earns citations on its own. Schema travels with citability the way a tailored suit travels with a job offer. Removing the suit does not revoke the qualifications, and putting one on does not create them.

Engine statements add a modest but genuine point in schema's favor. Bing's Fabrice Canel said in 2025 that schema helps large language models understand content for Copilot, and Bing's index matters beyond Copilot because ChatGPT's search experience launched on it. Google continues to say structured data helps its systems understand pages, even as its AI surfaces make no citation promises. No engine has ever said schema earns citations. Several have said it helps them read you. Those are different claims, and the second one is worth an hour of implementation time even if the first never materializes.

The two studies stop conflicting the moment you separate two jobs that schema gets asked to do. The first job is ranking: making an engine prefer your page over another page it has already understood. The Ahrefs data says schema does not do this job, and no engine claims it does. The second job is parsing: helping an engine work out, quickly and unambiguously, what your page contains, who published it, what your product costs, and which brand the pronouns refer to. The correlation data and the engine statements are consistent with schema doing this second job, and nothing in the Ahrefs study contradicts it, because already-cited pages have by definition already been parsed successfully.

Parsing help matters most exactly when you are unknown. An established publication with years of crawl history does not need JSON-LD for an engine to know its articles have authors and dates. A young company publishing a comparison page onto a domain the crawler met last month is a different case: every unambiguous signal about what the page is, who stands behind it and what the product costs reduces the chance of being misread or skipped. This is the population no schema study has isolated, and it is the population most Reachroller customers belong to, which is why our methodology treats schema as standard equipment on generated fix pages rather than as a promised lever.

There is also a cost asymmetry worth naming. Correct schema costs an hour to template and nothing to run. If the parsing benefit is real, you collect it. If it is not, you lost an hour. The tactics that cost real money, like rewriting your content strategy around structured data or buying tooling that promises schema-driven AI rankings, are the ones the evidence does not support. Cheap insurance, honest expectations. That is the whole position.

Understanding where schema could plausibly help requires knowing how the answer pipeline works. When an AI assistant answers a question with live retrieval, it searches an index, fetches a handful of candidate pages, extracts the passages that address the question, and synthesizes an answer with citations. Every stage before synthesis is machine reading at industrial speed, and the reading budget per page is small. OpenAI now operates its own crawler, OAI-SearchBot, and has roughly tripled its web crawl since August 2025 according to a Botify analysis. Perplexity maintains its own index, reported at more than 50 billion pages. Google's AI features cite from Google's organic index. None of these systems reads a page the way a person does; they parse it.

In that pipeline, structured data has one clear role: it is the part of your page that cannot be misread. A price inside a stylized pricing widget can be mangled by an extractor; a price in a Product schema field cannot. A brand name that collides with a dictionary word is ambiguous in prose; an Organization block with a sameAs array pointing at your official profiles resolves it. An article's publication date might live in a byline the template renders via JavaScript; the dateModified field is plain text in the source. Schema does not make an engine like your page. It lowers the chance the engine gets your facts wrong, and wrong facts in AI answers are their own expensive problem.

The precondition underneath all of this is being in the indexes at all. AI engines that retrieve read search indexes, so a page absent from Google and Bing does not exist for most of them, with or without markup. Submit new pages through Google Search Console and Bing Webmaster Tools before evaluating any on-page tactic, schema included. A perfectly marked-up page that no crawler has fetched is a tree falling in an empty forest.

Organization schema is the highest-leverage single block because brand disambiguation is the failure mode that hurts most in AI answers. If your company shares a name with a common word, a band, or another company, the sameAs array linking your site to your official LinkedIn, Crunchbase, GitHub and social profiles is the cheapest identity signal you can send. FAQPage schema earns its place on pages with genuine question-and-answer content, where it pairs naturally with the answer-first writing that works in AI search anyway; the full pattern is covered in our guide to FAQ pages that answer engines actually use.

Product schema matters for one specific reason: AI assistants state prices in answers, and they state whatever they parsed. A machine-readable price field, kept current, is your best defense against an assistant quoting the pricing you ran two redesigns ago. HowTo and the more exotic types sit at the bottom because Google has trimmed their rich results in classic search and no AI surface has given them a distinct role. Skip anything you would have to maintain by hand.

Use JSON-LD in a script tag rather than microdata woven through your HTML. Google recommends it, it keeps markup separate from presentation, and it survives redesigns because nobody has to re-thread attributes through new templates. Put the blocks in your page templates so they generate from the same data that renders the visible page. Hand-written schema drifts; templated schema stays true by construction.

The one rule that matters above all others: markup must describe what the page visibly says. Schema that claims content the visitor cannot see is treated as spam by Google's guidelines, and stuffing fields with keywords or fake ratings is the structured-data cousin of keyword stuffing, which the Princeton-led GEO study measured as performing near the bottom of all tactics for generative engines. Machines are precisely the audience you cannot fool twice: an engine that catches your markup lying about your page has a durable reason to distrust both.

Validate everything with Google's Rich Results Test or the Schema.org validator before shipping, because malformed JSON-LD is silently ignored and you get no error report from the engines. Then resist the urge to measure schema in isolation. Citation movement takes one to two weeks after indexing, AI answers are volatile run to run, and a single before-and-after check will tell you nothing trustworthy. If you want the honest measurement discipline, our guide to measuring AI visibility without lying to yourself covers repeated runs, stored answers and trend lines.

Perspective check. The only controlled academic study of generative engine optimization, the Princeton and Georgia Tech GEO paper presented at KDD 2024, tested nine on-page methods across thousands of queries and found the winners were content changes: adding quotations, adding statistics, and citing sources boosted visibility in generative engine responses by up to 40 percent, with the best methods improving roughly 22 percent on position-adjusted word count and 37 percent on subjective impression against baseline. Structured data was not among the winning levers. If you have one afternoon, spend it making your page more quotable, not more marked up. The full breakdown of those techniques is in our guide to writing content AI engines actually cite.

The second lever that outweighs schema is off your site entirely. Citation analyses consistently show AI engines leaning on independent sources, with 5W Research finding Wikipedia and Reddit alone accounting for over a quarter of ChatGPT citations in the U.S. Getting your brand named on a page an engine already trusts for your question often moves an answer faster than anything you publish yourself, marked up or otherwise. That is a digital PR problem, and no amount of JSON-LD substitutes for it.

Rank the work accordingly. Indexing first, because nothing else functions without it. Citable content second, because it carries the only measured on-page effect. Third-party mentions third, because engines weight independent sources heavily. Schema fourth, templated once and then left alone. Teams that invert this order spend their quarter perfecting markup on pages no engine retrieves, which is the most common failure pattern we see in new Reachroller accounts' first audits.

Here is how this playbook runs end to end, manually. Collect the twenty or so questions buyers actually ask in your category. Run them through the engines your buyers use and record which answers name you. For each question you lose, publish a page that answers it directly, with the statistics, quotations and cited sources the GEO research validated. Template your Article, Organization and FAQPage schema so every page carries it without per-page effort. Submit to Google Search Console and Bing Webmaster Tools. Wait one to two weeks, then re-run the questions several times each, because single runs mislead, and compare against your baseline.

Every step of that is doable by hand, and the honest cost is a working day to set up plus several hours a month to maintain. Reachroller automates the loop: it runs your buyer questions through ChatGPT today and Claude, Gemini, Perplexity and Grok as they roll out, scores only literal brand mentions in stored answer text you can open and read, and generates a publish-ready fix page for each question you lose, with the slug, title tag, meta description, schema markup and indexing steps included. The recheck then shows whether the answer flipped. Schema rides along in every fix because the parsing case is plausible and the cost is zero at generation time, and the recheck exists because plausible is where the evidence stops.

If you remember one thing from the conflicting studies, make it this: schema is a formatting decision, and AI visibility is won on substance. Format well, then put the effort where the measurements are.