Schema Markup for AI Search: What Actually Helps

Every few months a new claim goes round that one schema type is the secret to being cited by ChatGPT. It never is. Structured data is useful, cheap and worth doing properly, but its job is narrower than the hype: it makes facts about a page unambiguous to machines. This is what that buys you in AI search, which types are worth shipping, and which claims to ignore.

Schema markup is a vocabulary, defined at schema.org, for stating facts about a page in a format machines can parse without guessing. On most sites it ships as JSON-LD, a block of structured data in the page's HTML that says, in effect: this is an article, this person wrote it, it was published on this date and updated on that one, and it belongs to this organization.

The useful word there is unambiguous. A human reading your page can work out that "Updated June" near the byline is a date and that the logo in the header is your company. A machine can usually work it out too, but it is inferring. Schema removes the inference. That is all it does, and it is why the honest case for it is accuracy rather than ranking.

It is also why schema cannot rescue a page whose answer is not in the HTML to begin with. If the content a crawler receives is an empty shell that fills in with JavaScript, markup describing it is describing nothing. Our piece on answers that are not in the HTML covers that more fundamental problem, and it is the one to fix first.

It helps to separate what the platforms have said from what people have inferred.

Google: structured data feeds search features such as rich results, and helps Google understand page content. Google has also said that no special markup is required to appear in AI Overviews: the same eligibility rules as ordinary search apply. Since AI Overviews are built on Google's index, anything that helps Google understand a page is at least not wasted there.

Bing: Microsoft has said publicly that structured data helps its systems understand content, and Bing's index sits behind Copilot answers. That is a stronger statement than most, but it is still not a promise that markup earns a citation.

ChatGPT, Perplexity, Claude and the rest: none has published how its retrieval pipeline treats JSON-LD. Many pipelines convert a page to plain text before a model reads it, and whether structured data survives that step is not documented. Anyone telling you schema is a direct ranking factor in these engines is guessing.

These are the types that describe facts an answer engine needs to get right about you. In priority order, roughly:

Organization is the one most sites get wrong. The valuable part is sameAs: links to your LinkedIn page, Crunchbase profile, GitHub, Wikipedia entry if you have one, and any other profile that is unambiguously you. That is what ties your site to the same entity everywhere else on the web, and entity confusion (an answer engine mixing you up with a similarly named company) is a real, common failure in AI answers.

Dates are the second. Answer engines lean towards fresh sources for anything time-sensitive. If your dateModified is missing, or is set to the build date on every deploy whether the content changed or not, you are either hiding freshness or faking it. Set it when the content actually changes.

FAQPage is fine, in its place. A real question and answer section, marked up, is a clean structure that mirrors how people ask AI tools questions. Google stopped showing FAQ rich results for most sites in 2023, so the visible reward in classic search is gone. Keep it on pages with genuine FAQs and do not invent FAQs to have something to mark up.

"Schema gets you cited by ChatGPT." There is no published evidence for this. Pages get cited because they are retrievable, relevant and say something specific. Markup on a thin page is still a thin page.

"Speakable is how you get into voice and AI answers." Speakable was introduced by Google as a beta for news content read aloud by Google Assistant, limited in scope. It is harmless and some sites, including this one, use it to point at the answer passage, but it is not a general AI visibility switch.

"More types is better." Marking up every possible entity on a page adds maintenance and risk without a demonstrated benefit. HowTo rich results, for example, were removed from Google search in 2023. Ship the types that describe true, important facts and stop there.

"llms.txt replaces schema" or the reverse. They do different jobs. Schema describes a page in place. An llms.txt file is a proposed, site-level map that points language models at your most useful content in plain markdown. Neither is an official ranking signal for any major engine. Both are cheap, and you can do both.

Render it on the server. Put JSON-LD in the HTML the server sends, not in a script that runs after load. Google can render JavaScript, but many AI crawlers fetch the raw HTML and move on. In a framework like Next.js that means generating the JSON-LD in the page component and emitting it as a script tag at render time, which is how this site does it.

Generate it from the same data as the page. If the headline, dates and FAQ answers in the markup come from the same source as the visible text, they cannot drift apart. Hand-written JSON-LD pasted into a CMS field is how pages end up with a price in the markup that changed a year ago.

Use stable IDs. Give your Organization and authors an @id and reference it from every article, rather than repeating a slightly different description on each page. One entity, described once, linked everywhere.

Validate twice. Google's Rich Results Test tells you what Google can use for search features, and the Schema Markup Validator at validator.schema.org checks the markup against the vocabulary itself. A page can pass one and fail the other, so run both. Our free Domain Score Checker also flags missing structured data alongside AI crawler access.

Then stop. Once the core types are accurate and rendered server-side, more markup is unlikely to move anything. Your time is better spent on what actually earns citations: pages that answer a specific question in their first lines, with something an answer engine cannot get elsewhere. If you want a starting structure for that, the AEO FAQ page template is a practical one.