Schema markup for AI search: what Google documents and what is only claimed

Does schema markup for AI search make a difference?

For Google, the documented answer is no: there is no special markup for its AI features. Google says so in plain words on the page it wrote for site owners about AI Overviews and AI Mode.

Google's page AI features and your website says: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add." The same page says: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." It then points site owners back to ordinary SEO best practices.

That is not the same as saying structured data is useless. The same Google page lists, among its general best practices, "Making sure your structured data matches the visible text on the page". Structured data is a label you add to a page so a machine knows what each part is. Google uses it in ordinary search, and the rest of this guide sorts what is documented from what is only claimed.

One word on names. Schema markup, structured data and schema.org all point at the same idea here: a shared vocabulary from schema.org, written into the page, usually as a JSON-LD block. JSON-LD is a small piece of code in a script tag that describes the page in a fixed format.

What Google actually documents

Google documents five things that matter here, and none of them is a ranking boost in AI answers. Each row below quotes or closely follows a Google page read on 10 October 2026.

The FAQ change is worth dwelling on, because FAQ markup is often sold as an AI search tactic. Google first limited FAQ rich results in 2023 to "well-known, authoritative government and health websites", and its documentation updates log records that the feature stopped appearing on 7 May 2026. HowTo rich results were removed earlier, in 2023. The markup does no harm, but the visible reason to add it in Google Search no longer exists.

What is claimed elsewhere, and how strong the evidence is

Outside Google's documentation, the evidence is mostly statements and vendor claims. Some of it may be true. Very little of it can be checked.

The Microsoft statement is the strongest of the vendor claims, because it comes from someone who works on Bing. It is still a reported remark, not a documented rule. It says markup helps a model understand content. It does not say markup makes a page more likely to be cited, and it says nothing about ChatGPT, Perplexity or Gemini.

The research that does exist points at the visible page. The GEO paper tested nine ways to edit a page, and its best methods were adding citations, quotations and statistics to the text. None of the nine involved structured data. That does not prove markup has no effect. It means the one published, peer-reviewed benchmark in this area did not test it.

Google documents that no special schema is needed for its AI features. Anyone claiming otherwise should be able to show you the study.

Which schema types are worth adding

Add the few types that describe what your page plainly is, and stop there. They are cheap, Google documents a use for most of them, and they cannot mislead if they match the page.

Other assistants may read the same markup when they fetch a page, or they may read only the visible text. We found no page from OpenAI, Perplexity or Google saying their assistants reward schema markup in their answers. Since the markup has to match the visible text anyway, the visible text is the part that has to be right.

A note on llms.txt and other AI files

The same advice covers the new files sold as AI search fixes. Google's AI features page says you do not need to create "new machine readable files, AI text files, or markup" to appear in its AI features. Its documentation updates log for June 2026 adds that llms.txt files are not needed for Google Search and will not help or hurt your visibility or rankings there.

That log entry also says it is fine to keep such files for other services that use them. So the honest position is narrow. For Google, they do nothing. For other assistants, nobody we read has published evidence either way, and the cost of a small file is low.

Rules that apply to any markup you add

Markup must describe what a reader can see, and it must not mislead. Google's general structured data guidelines are short on this and strict.

• Do not mark up content that is not visible to readers. Google's example is that if the JSON-LD describes a performer, the page body must describe the same performer.
• Do not mark up irrelevant or misleading content, such as fake reviews. Google says not to use structured data to deceive or mislead users.
• A structured data manual action costs a page its eligibility for rich results. Google says it does not affect how the page ranks in web search.
• Correct syntax is not enough. Google says markup that breaks its quality rules may not be shown, or may be marked as spam.

There is a practical reason behind these rules for AI search too. If your markup says one price and your page says another, a reader cannot tell which is true, and neither can a model reading both. Keep one source for each fact, and generate the markup from it where you can.

A sensible order of work

Do the things Google documents as requirements first, then the cheap markup, then everything else. This order puts effort where the evidence is.

1. Make sure the page can be crawled and indexed. Google says a page must be indexed and eligible for a snippet to appear in its AI features. Other assistants have their own crawlers, which your robots.txt can allow or refuse, as our robots.txt guide explains.
2. Write the visible answer well. Put a direct answer under each heading, give figures with their sources, and use tables for comparisons.
3. Add Organization markup to the home page, with your name, logo and official profiles.
4. Add Article or BlogPosting markup to articles, and Product markup to product pages with a public price.
5. Check every marked-up fact against the visible text. Remove any markup that describes something the page does not show.
6. Test with Google's own tools, then leave it. Markup is maintenance, not a growth lever.

How to tell whether it changed anything

Measure the AI answers themselves, before and after, on the same questions. Adding markup and watching traffic will not tell you much, because many other things change in the same weeks.

Pick the questions your buyers ask, ask each assistant the same set more than once, and record whether you are named and which pages are cited. Answers vary from one run to the next, so a single run is not a result. Agentcites does this on a schedule: each week it asks ChatGPT, Perplexity, Gemini and DeepSeek the same buyer questions and records which brands they name and which pages they cite. It does not check your structured data. Our method page explains how answers are counted, and the features page lists what is measured.