Does schema markup get you cited by AI?

Probably not, on its own. The one controlled study on this, Ahrefs in May 2026, found AI Overview citations fell 4.6% on pages that added JSON-LD, a small but statistically significant drop, with no measurable movement on AI Mode or ChatGPT. Schema still earns its place for conventional search features.

Almost every AEO guide published in the last two years opens with "add structured data." That advice has one controlled test behind it and the test came back negative. Desses sells an AEO audit that covers schema. Selling you a schema project would be good for us. I'd rather you put the budget where something moves.

Schema markup is a vocabulary from schema.org for labelling what the things on a page are, usually written as JSON-LD: a block of JSON in the head or body of the HTML that describes the page's entities to a machine and shows nothing to a reader. That last clause is the whole story of this article.

What the one controlled study found

Ahrefs published it on 11 May 2026: 1,885 pages that added JSON-LD between August 2025 and March 2026, matched against a control group of exactly 4,000 pages that didn't, compared four separate statistical ways.

AI Overview citations came out at −4.6%, and that one held up as statistically significant. AI Mode landed at +2.4% and ChatGPT at +2.2%, both close enough to zero to be noise.

Then the authors put the knife in their own study, and nobody quoting the headline number bothers to carry this part. Every page in the sample was already receiving 100 or more AI Overview citations before any schema went on it. The study watched pages already inside the citation pool and measured what happened when they added markup. It has nothing to say about a page trying to get into that pool from zero, which is where most people reading a schema guide are.

The claim that survives is narrow. On pages already being cited, adding JSON-LD failed to help, measured against a control, four ways. Nobody has shown it does damage. Everything else circulating on this topic is a case study of one site with no control and a happy ending.

Somebody tested the mechanism directly

searchVIU published a test on 2 December 2025: one page carrying eight product prices, placed five different ways, in visible HTML, in JavaScript-rendered content, in hidden JSON-LD, in Microdata and in RDFa, then put to five AI systems. Not one of the five pulled a single price out of the JSON-LD on a direct page fetch.

Read that middle column carefully, because the line most people repeat from this test is wrong. Gemini was the only system that ran JavaScript when it fetched the page live. Google AI Mode and Perplexity both reached the JavaScript-rendered price too, after the page had been indexed, which means their crawlers execute scripts and their live readers don't.

One page, one run, so treat it as directional. A crawler that reads raw HTML and doesn't run scripts is reading your visible text. Your JSON-LD sits in the same file, unread.

Google and Bing openly disagree

Google's Search Central documentation on AI features says it plainly: "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."

Microsoft's Fabrice Canel confirmed at SMX Munich in March 2025 that Bing and Copilot's LLMs use schema markup. That reached the record through David Mihm's LinkedIn summary of the session, which Canel confirmed in a comment. No schema types were named, and Microsoft has published nothing since that narrows it down.

Both positions can hold at once. Schema gets parsed at index-build time by a full search crawler, and ignored at live-fetch time by a model reading raw HTML in the moment. Copilot is grounded in the Bing index, so Canel is describing what Bingbot does. searchVIU describes what a chatbot does mid-conversation. It also explains why AI Mode and Perplexity saw the JavaScript price after indexing and missed it live.

It's a mechanism that fits the evidence. Nobody at Google or Microsoft has confirmed it, and I'm not going to present a hypothesis as a finding.

FAQPage is dead twice over

This one has hard dates, and most guides still recommending it haven't checked any of them.

Google cut FAQ rich results back on 8 August 2023, restricting them to "well-known, authoritative government and health websites." That was the first death. The feature stopped appearing entirely on 7 May 2026, and Google removed the documentation on 15 June 2026. Search Console API support ends in August 2026. HowTo went the same way, and Practice Problem was deprecated in January 2026.

The surface is gone. No FAQPage markup you write today renders anywhere in Google.

That leaves the fallback you'll see in every comment thread: the rich result died but FAQ schema still earns AI citations. No provider has said it. No study has measured it. It survives because a lot of people have FAQ blocks they don't want to have wasted.

The Q and A format does carry. In visible HTML, on the page, where a person reads it. Keep the format. Drop the markup.

Everything has to be in the visible text

Anything you need an AI system to state about you has to exist in visible body text on the page. Markup supplements that text.

Google's one structured-data instruction in the context of AI features is a validity instruction: make sure your structured data matches the visible text on the page. Read that next to the extraction tests and it stops being a compliance note. Schema describing content the page doesn't display is both a policy violation and invisible to the systems you were writing it for. Two failures, one cause.

If your product's price only exists in an Offer block, no system in the searchVIU test could see it on a live fetch. Put the price on the page in a sentence. Then mark it up as well.

What's still worth adding

Every type on that list earns its place in the conventional search results that still send you almost all of your traffic. AI citations are a separate question, and none of these types has been shown to move them.

Where this argument stops

Read the top of this page as permission to strip JSON-LD off your site and you've read it wrong. Product and Offer are doing measurable work in Google Shopping right now. Article and BreadcrumbList still render. Organization schema costs an hour once and never needs touching again. Trading a working search feature for a shot at a channel that sends 0.14% of web visits is a bad trade in both directions.

One group this article doesn't apply to: anyone running no structured data at all. Add Organization, Article and BreadcrumbList once, correctly, matching what's on the page, and stop. Don't build a schema programme. Don't buy a schema tool. Don't let anyone sell you a quarterly structured-data roadmap on the promise of AI visibility. That promise has one controlled test behind it and the test came back at −4.6% on pages that were already being cited.

What we'd build, on Framer

Framer ships no schema natively. Every type is custom code, in Project Settings, and CMS variables interpolate straight into the JSON with {{FieldName | json}} escaping safely. The full build is in the Framer guide.

On a typical client build that's one Organization block with sameAs sitewide, Article on the insights CMS template, BreadcrumbList on the nested routes. Half a day, once, then it maintains itself off the CMS fields.

The other half goes where the citations do. Framer pre-renders every page to HTML on its servers, so a crawler that never runs JavaScript still receives the full text. That text is the asset. If the answer to the question in your H1 isn't in the first paragraph of it, no amount of markup will put it there. More on that in our Framer development work.

If you want to know what your site is shipping, send me the URL. I'll send back what's in your head tag, what a non-rendering crawler sees in your visible HTML, and which of it is doing nothing.