Schema markup for AI search: what the 2026 data shows

Schema markup does not appear to buy you AI citations. Two separate 2026 studies, one from Ahrefs and one peer-reviewable paper on SSRN, measured it directly and found no meaningful lift in how often ChatGPT, Perplexity or Google’s AI surfaces cite a page. Ship structured data anyway, because one AI vendor has confirmed it uses it and classic search still runs on it, but do not sell it, buy it, or budget it as a generative-search lever.

Does schema markup increase AI citations?

No, not on the evidence published so far. Ahrefs ran a difference-in-differences study on 1,885 pages that added schema between August 2025 and March 2026, against roughly 4,000 control pages, and published the results on 11 May 2026. Google AI Overviews citations fell 4.6% for the pages that added markup, a statistically significant decline. Google AI Mode moved 2.4% and ChatGPT moved 2.2%, both inside the noise band.

A cross-platform study by Kurt Fischman, posted to SSRN on 20 February 2026, came at it from the other direction: take 1,006 pages and 730 actual AI citations, then check what the cited pages have in common. Schema was present on 43.1% of cited pages and 44.8% of non-cited pages. Statistically indistinguishable.

What did predict citation in that dataset was Google organic rank, at an odds ratio of 0.762 per position. Pages at position 1 were cited about 43% of the time. By position 7 that had fallen to roughly 5%. Rank is doing the work that structured data is often credited with.

FAQPage markup is being switched off, not read harder

Google added a deprecation notice to its FAQ structured data documentation in May 2024, after restricting FAQ rich results in August 2023 to what it called well-known, authoritative government and health websites. Full removal of the feature landed on 7 May 2026, and Search Console API support for the report ends this month, August 2026.

That deprecation is scoped to classic rich results. Google has published nothing tying FAQPage markup to AI Overviews either way. So the honest position on FAQPage in August 2026 is that its one documented benefit has already been switched off, and no vendor has documented a replacement benefit on the AI side.

Keeping FAQPage in your templates is still defensible, since the data behind it is usually the same content collection that renders the visible FAQ block, and the marginal cost of emitting it is zero. Rebuilding your content strategy around it is not.

Which AI surfaces have actually said they read schema

One. Microsoft’s Fabrice Canel said at SMX Munich on 20 March 2025, as reported by Search Engine Land, that schema markup helps Microsoft’s LLMs understand your content. That remains the only on-the-record confirmation from a major AI search vendor.

Surface	Official statement on reading schema	Date
Bing and Copilot	Confirmed: schema helps Microsoft’s LLMs understand content	20 March 2025
ChatGPT and OAI-SearchBot	None. OpenAI’s crawler documentation covers robots.txt controls and never mentions JSON-LD or schema.org	checked August 2026
Perplexity	None. Its published structured-output docs cover developer API responses, not crawler-side parsing	checked August 2026
Google AI Overviews and AI Mode	None either way on AI surfaces specifically	as of August 2026

Anyone quoting you a citation-rate percentage for schema on ChatGPT or Perplexity is quoting an SEO blog, not a vendor. We could not trace those figures to a primary source, and we are not going to repeat them here.

What the studies do not settle

Neither 2026 study separates schema types. Both measure schema presence in aggregate, so nobody has published evidence that Service and Offer markup performs differently from Article or Organization markup on any AI surface. If someone tells you Offer markup specifically is what gets you into an AI shopping answer, ask which study, then check whether it separated types. So far none of them do.

The Ahrefs result also measures adding schema to pages that were already being crawled and cited. It does not tell you what happens to a page that has no structured data at all, on a site with no Organization markup, competing in a category where every rival has both. Absence of a proven lift is not proof that stripping everything out is safe.

Both studies are also snapshots of systems that change monthly. The reasonable reading is that as of mid-2026 structured data is not a citation lever, not that it can never become one.

What we ship on every build, and why we keep shipping it

Every page on this site carries server-rendered JSON-LD, written into the static HTML at build time. Nothing is injected by client-side JavaScript, because a crawler that does not run scripts should still get the full picture.

Sitewide we emit Organization, with a nested ContactPoint, and WebSite. Service pages emit Service with a nested Offer carrying the real price. Blog posts and case studies emit Article with a Person author. The pricing page emits an OfferCatalog. FAQPage runs on the homepage, the method page, the pricing page and every service page.

The part that matters more than the type list: on service pages the FAQ schema and the visible FAQ block read from the same faqs array in the content collection, and the Service schema reads the same fields the visible page renders. One source, two outputs. Structured data that quietly contradicts the page it sits on is worse than none, and the only reliable way to prevent that is to generate both from the same data at build time rather than maintaining a parallel copy by hand.

We keep it for three reasons, none of them AI citations. Bing and Copilot have said they use it. Classic Google rich results still run on it. And it costs us nothing per post, because it regenerates from the same collections that render the pages, the same way our sitemap, RSS feed and llms.txt do. We wrote up that build-time approach in our piece on llms.txt, where the evidence landed in a similar place: cheap to keep current, not a lever to plan around.

Where to put the effort instead

Rank is the variable with published evidence behind it, and rank comes from the page. Answer the target query directly in the first two or three sentences. Give each H2 a self-contained answer near the top, so a model quoting that section alone still says something true and complete. Name a real statistic, name who published it, and date it in the sentence. Put all of that in server-rendered HTML.

That is the mechanics we set out in our guide to generative engine optimization, and it is what the citation research keeps pointing back to. The markup describes the page. It does not rescue one.

The other half is cadence, which is where most teams actually stall. Our own group runs a 363-product catalogue with daily publishing, structured data on every page, and a human approving every merge. The structured data is table stakes there. The publishing rhythm is what moves the numbers. When we build a content engine for a client, both ship together at a fixed price, and the client owns the repo, the pipeline and the schema helpers outright at handover.

The plain version

Emit clean structured data because it is nearly free and one vendor has confirmed it uses it. Generate it from the same source as your visible content so the two can never drift. Do not pay anyone a premium for a schema retrofit sold as an AI-citation strategy, because as of August 2026 the two studies that measured that directly found a decline on one surface and noise on the rest.

Describe your current setup and you will get a written plan back within one business day: what to emit, what to retire now that FAQ rich results are gone, and what to publish instead. No meetings, fixed price, and you own the pipeline. Start async.