Which Schema Types Actually Help With AI Answers?

Structured data

Machine-readable markup, usually JSON-LD in the page head, that describes what a page is about using the schema.org vocabulary.

Also known as: Schema markup, JSON-LD, Rich result markup

The eligibility rule Google publishes for AI Overviews and AI Mode mentions no markup at all. A page has to be indexed and eligible to be shown in Google Search with a snippet, and Google adds that there are no additional requirements and no other special optimisations necessary. Markup is not one of the gates.

None of the other engines contradict that. OpenAI, Anthropic, Perplexity and Microsoft document their crawlers and their retrieval behaviour, and none of them document a schema type that changes source selection. Any claim that a specific markup type is weighted during selection is an inference nobody has published evidence for.

The right frame is that markup is cheap insurance, not a lever. It costs almost nothing to emit correctly, it powers a handful of surviving Google rich results, and it is semantically honest. It is not the thing standing between a page and a citation.

The FAQ case is worth understanding rather than just accepting, because it explains the whole category. What correlated with citation in the Semrush data was the visible question and answer text on the page, at plus 25.5%, not the JSON-LD describing it. The markup was always a wrapper around the thing that mattered, and when Google switched the rich result off, the wrapper lost its last measurable job.

The evidence genuinely conflicts, so it is worth laying out both sides. Semrush, across 11,882 prompts and 304,805 cited URLs, found structured data elements associated with citation at plus 21.6%. The GEO-16 analysis by Kumar and Palkhouski, covering 1,702 citations across 16 B2B SaaS verticals, reported structured data at r=0.63 and semantic HTML at r=0.65.

Both are observational. Against them sits the SIGIR 2026 paired test by Vishwakarma, Kumar and Jamidar, which changed one factor at a time across 252,000 trials and six models, and found formatting and structure changes negligible. And Citera, analysing roughly 350,000 B2B SaaS articles, found schema markup present on 69 to 72% of articles at every ranking position, with no gradient whatsoever.

A confound explains all four results at once. Sites that emit clean markup tend to be sites with a competent publishing setup, which also produces clearer writing, better dates and more original content. The correlation is real and the causal arrow points at the publisher, not the markup.

Dataset is the interesting one. It is a supported Google feature through Dataset Search, it takes a name, a description, a temporal coverage, a creator and a licence, and it is the honest description of any page that publishes a measured aggregate. Marketing sites almost never emit it, because marketing sites almost never publish measured aggregates.

Oppira emits a Dataset node on the pages that publish our own aggregate competitor measurements, with the sample size and the as-of date stated in the visible text next to every figure. To be clear about what that buys: it makes the page eligible for Dataset Search and it is semantically true. There is no published evidence that Dataset markup increases AI citations, and anyone claiming otherwise is guessing.

This rule quietly invalidates a whole family of clever tactics. Stuffing extra questions into FAQPage markup that do not appear on the page, emitting a fresher dateModified than the visible date, or listing an author who has no byline are all violations of it, and all three are common.

There is a second reason to follow it beyond compliance. Engines quote rendered text, and the Semrush factors that associated positively with citation, clarity and summarisation at plus 32.8% and visible question-and-answer format at plus 25.5%, are properties of the text a human sees. Markup describing content that does not exist cannot be quoted, because there is nothing there to quote.

So treat JSON-LD as a mirror. Anything in it should be a machine-readable restatement of something already on the page, generated from the same source, which also means it cannot drift out of sync when the page is edited.

What you will not find is a report telling you whether markup helped an AI answer. No engine exposes that, so resist the temptation to infer one from a citation that arrived the same week you shipped some JSON-LD.